[
  {
    "title": "Anthropic muss Fable 5 und Mythos 5 nach US-Regierungsanordnung abschalten",
    "summary": "Die US-Regierung hat Anthropic per Exportkontrolldirektive gezwungen, den Zugang zu den erst kürzlich veröffentlichten Modellen Fable 5 und Mythos 5 vollständig zu sperren. Das US-Handelsministerium befürchtet, dass ein Jailbreak die Sicherheitsfilter von Fable 5 für Cybersicherheit, Chemie und Biologie umgehen könnte und damit eine Gefahr für die nationale Sicherheit darstellt. Anthropic widerspricht der Einschätzung und argumentiert, dass der gemeldete Jailbreak nur 'geringfügige' Schwachstellen findet und andere öffentliche Modelle wie GPT-5.5 ähnliche Fähigkeiten besitzen. Das Unternehmen entschuldigt sich bei Kunden für die abrupte Unterbrechung und kündigt weitere Details innerhalb von 24 Stunden an.",
    "source_url": "https://arstechnica.com/ai/2026/06/anthropic-shuts-down-fable-mythos-models-following-trump-admin-directive/",
    "tag": "tag-ki",
    "full_text": "Anthropic completely shut off access to its Mythos 5 and Fable 5 models Friday night, just days after they were launched. The move comes after Anthropic’s receipt of a US Commerce Department directive Friday evening, subjecting the new models to export controls restricting their use anywhere outside the United States. In a message posted Friday night , Anthropic said the only way for it to ensure compliance with that government order in the immediate term “is that we must abruptly disable Fable 5 and Mythos 5 for all our customers.” Access to other Anthropic models is not affected. An Axios report cited an administration official saying that the administration is concerned by reports of a jailbreak that reportedly gets around broad classifier-based safeguards meant to block Fable 5 prompts regarding cybersecurity, chemistry, and biology. The administration reportedly requested a pause in the release of these models to gain time for the “national security apparatus” to be “hardened” against this kind of threat. That hardening could be complete “in the next few weeks,” Axios’ source suggested. In its Friday night announcement post, Anthropic said the government has only provided it with “verbal evidence of a potential narrow, non-universal jailbreak” that involves getting Fable 5 to review a specific codebase for software flaws. The company says it has only seen evidence of this kind of jailbreak being used to find “minor” and “relatively simple” software vulnerabilities, and that other publicly available models like GPT-5.5 has similar capabilities on this score. “We are complying with the government’s legal directive and are removing access to Fable 5 and Mythos 5 for all users,” Anthropic writes. “However, we disagree that the finding of a narrow potential jailbreak should be cause for recalling a commercial model deployed to hundreds of millions of people. If this standard was applied across the industry, we believe it would essentially halt all new model deployments for all frontier model providers.” Earlier this month, President Trump signed an executive order urging AI model makers to submit to voluntary government security testing. That order came after an initial signing ceremony planned for last month was abruptly postponed amid reported concerns of disagreements about it within the administration. Anthropic apologized to customers for a “disruption” that it said is the result of a “misunderstanding,” and said it will release more details about the situation in the next 24 hours."
  },
  {
    "title": "Anthropic rudert bei Fable 5 zurück: Verdeckte Sicherheitseingriffe werden sichtbar",
    "summary": "Nach massiver Kritik aus der Entwickler-Community macht Anthropic die bisher versteckten Schutzmechanismen seines neuen KI-Modells Fable 5 transparent. Statt Antworten heimlich zu manipulieren, fällt das Modell bei erkannten Distillation-Versuchen nun sichtbar auf Claude Opus 4.8 zurück. Anthropic entschuldigt sich ausdrücklich für die ursprüngliche Entscheidung zugunsten verdeckter Eingriffe und räumt ein, dass Nutzer nachvollziehen können müssen, welche Schutzmaßnahmen aktiv sind. Die Umstellung führt vorübergehend zu mehr Fehlalarmen, da die Klassifikatoren konservativer arbeiten müssen.",
    "source_url": "https://www.heise.de/news/Fable-5-Anthropic-stoppt-verdeckte-Eingriffe-11330094.html",
    "tag": "tag-ki",
    "full_text": "Anthropic reagiert auf die Kritik an den Schutzmechanismen seines neuen KI-Modells Fable 5 . Das Unternehmen will umstrittene, verborgene Sicherheitsmaßnahmen künftig sichtbar machen und entschuldigt sich ausdrücklich für deren bisherige Umsetzung. Konkret geht es um Schutzmechanismen gegen sogenanntes Distillation – also den Versuch, die Ausgaben eines leistungsfähigen Sprachmodells zum Training konkurrierender KI-Systeme zu nutzen. Weiterlesen nach der Anzeige Die Kontroverse entzündete sich an einem Schutzverhalten von Fable 5, bei dem das Modell verdeckt auf Distillation-Anfragen reagierte. Anthropic sah ursprünglich einen unsichtbaren Mechanismus vor, der solche Versuche zur Modellentwicklung im Hintergrund erkennt und die Antworten gezielt verändert oder verschlechtert. Die Nutzer sollten davon nichts mitbekommen. Forscher und Entwickler kritisierten das als intransparent und warnten, dass solche verdeckten Eingriffe auch Tests und wissenschaftliche Untersuchungen des Modells verfälschen. Fable 5 fällt künftig sichtbar auf Opus 4.8 zurück In einem Beitrag auf X kündigt Anthropic nun eine Kurskorrektur an. Künftig behandelt das Unternehmen erkannte Distillation-Anfragen sichtbar. Statt Antworten heimlich zu verändern, fällt Fable 5 in solchen Fällen auf das ältere Modell Claude Opus 4.8 zurück – genau wie es bereits bei den Schutzmaßnahmen für Cybersecurity und Biologie der Fall ist. Die Nutzer sollen dabei jedes Mal einen entsprechenden Hinweis sehen. Für API-Kunden will Anthropic zudem den Grund einer Ablehnung explizit zurückgeben. Ein serverseitiger Fallback für API-Anfragen soll in den kommenden Tagen folgen. Damit lässt sich künftig erkennen, ob eine Antwort von Fable 5 oder vom Fallback-Modell stammt. Anthropic räumt die falsche Abwägung ein Das Unternehmen gibt zu, mit dem ursprünglichen Ansatz falsch gelegen zu haben. Sichtbare Schutzmechanismen lassen sich zwar leichter analysieren und gezielt umgehen, weshalb ihre Absicherung mehr Zeit kostet. Unsichtbare Schutzmaßnahmen lassen sich dagegen enger auf bestimmte Szenarien zuschneiden und verursachen weniger Fehlalarme. Aus diesem Grund habe man sich zunächst für den verdeckten Ansatz entschieden, um Fable 5 schnell und sicher bereitzustellen. Lesen Sie auch Wirklich zu gefährlich? | c't 3003 c't Magazin Fable 5 im Test: Das kann das teuerste Anthropic-Modell c't Magazin Fable 5 blockiert auch sicheren Code iX Magazin Das wird teuer: Anthropics Claude Mythos 5 erscheint als Fable 5 mit Schranken Reaktion auf Claude Mythos? Bundesregierung gründet KI-Sicherheitsinstitut iX Magazin Mehr anzeigen Weniger anzeigen Rückblickend sei das die falsche Entscheidung gewesen, schreibt Anthropic. Die Nutzer sollten nachvollziehen können, welche Schutzmaßnahmen aktiv sind und warum. Dafür entschuldigt sich das Unternehmen ausdrücklich. Weiterlesen nach der Anzeige Mehr Transparenz, vorerst mehr Fehlalarme Die Umstellung hat allerdings Nebenwirkungen. Um die Systeme trotzdem vor Jailbreaks abzusichern, müssen die zugrunde liegenden Klassifikatoren zunächst konservativer arbeiten. Das führt vorübergehend zu mehr Fehlklassifikationen. Solche False Positives entstehen, wenn das Modell harmlose Anfragen fälschlich als riskant einstuft. Genau hier setzt ein Großteil der bisherigen Kritik an. Videos by heise mehr Videos c't 3003 heise & ct Peertube Kritik aus der Sicherheitscommunity Die Ankündigung folgt nur wenige Tage auf heftige Kritik von Sicherheitsforschern an Fable 5. Mehrere Experten beklagen, dass die Cybersecurity-Schranken des Modells nicht nur brisante Anfragen erfassen, sondern auch alltägliche Aufgaben aus Softwareentwicklung und IT-Sicherheit. Genannt wurden unter anderem Code Reviews, das Schreiben sicheren Codes, Schwachstellenanalysen, Incident Response oder schlicht das Lesen sicherheitsrelevanter Fachartikel. Fable 5 ist die öffentlich verfügbare Variante von Anthropics neuem Spitzenmodell Mythos 5. Letzteres bringt keine vorgeschalteten Schutzmechanismen für Cybersecurity, Biologie, Chemie und Distillation mit. Anthropic justiert die Cyber- und Bio-Filter nach In seiner Stellungnahme verspricht Anthropic auch Änderungen an den Cyber- und Bio-Safeguards. Die entsprechenden Klassifikatoren stelle man derzeit so ein, dass sie seltener bei harmlosen Anfragen anschlagen. Nutzer, die eine Fehlklassifikation vermuten, sollen diese melden – über Feedback-Funktionen in Claude Code und Claude.ai sowie über ein Einspruchsformular für API-Anfragen. Ob die Anpassungen ausreichen, bleibt abzuwarten. An den Schutzmaßnahmen selbst hält Anthropic ausdrücklich fest – diese hatten die Kritiker allerdings auch nicht infrage gestellt. ( fo )"
  },
  {
    "title": "Bundestag beschließt KI-Gesetz: Bundesnetzagentur wird zentrale KI-Aufsichtsbehörde",
    "summary": "Der Deutsche Bundestag hat das nationale Umsetzungsgesetz für den EU AI Act verabschiedet und die Bundesnetzagentur als zentrale Marktüberwachungsbehörde für Künstliche Intelligenz bestimmt. Die Behörde wird Beschwerdestelle für Bürger, Beratungsstelle für Unternehmen und Betreiber von KI-Reallaboren. Wirtschaftsverbände begrüßen die Rechtssicherheit, während Opposition und Zivilgesellschaft Defizite beim Grundrechtsschutz beklagen. Ein verpflichtendes Transparenzregister für KI-Einsatz in Behörden wurde abgelehnt. Die jährlichen Kosten werden auf 15,9 Millionen Euro für den Bund geschätzt.",
    "source_url": "https://www.heise.de/news/Bundestag-beschliesst-KI-Gesetz-Bundesnetzagentur-wird-zentrale-Aufsicht-11330801.html",
    "tag": "tag-ki",
    "full_text": "Kurz vor 22 Uhr hat der Bundestag am Donnerstag das nationale Umsetzungsgesetz für die KI-Verordnung der EU verabschiedet und die Bundesnetzagentur als zentrale Marktüberwachungsbehörde für Künstliche Intelligenz in Deutschland bestimmt. Wirtschaftsverbände begrüßen die gewonnene Rechtssicherheit. Opposition und Zivilgesellschaft beklagen dagegen Defizite beim Grundrechtsschutz und warnen vor einem föderalen Kompetenz-Wirrwarr. Weiterlesen nach der Anzeige Mit den Stimmen der Koalition von CDU/CSU und SPD beschloss das Parlament den vom Digitalausschuss noch geänderten Gesetzentwurf . AfD, Grüne und Linke votierten dagegen. Zugleich nahmen die Abgeordneten eine Entschließung an, die einen innovationsfreundlichen und möglichst bürokratiearmen Rahmen für KI-Anwendungen schaffen soll. Oppositionsanträge für ein Verbot biometrischer Fernidentifikation im öffentlichen Raum sowie ein verpflichtendes Transparenzregister fanden keine Mehrheit. Mit der Initiative schafft Deutschland knapp zwei Jahre nach Inkrafttreten des AI Acts die nötigen nationalen Aufsichts- und Bußgeldstrukturen. Die Bundesnetzagentur übernimmt dabei die Marktaufsicht , sofern keine spezialisierten Fachbehörden zuständig sind. Bei der Behörde soll zudem ein Koordinierungs- und Kompetenzzentrum als Ansprechpartner für europäische Institutionen und zur Bündelung nationaler Expertise entstehen. Bundesnetzagentur erhält Schlüsselrolle Für die Bürger wird die Bundesnetzagentur zur Beschwerdestelle bei vermuteten Verstößen gegen KI-Vorgaben. Gleichzeitig soll sie Unternehmen beraten und mindestens ein KI-Reallabor betreiben, in dem neue Anwendungen vorab getestet werden können. So will der Gesetzgeber Start-ups und mittelständische Unternehmen unterstützen. Videos by heise mehr Videos c't 3003 heise & ct Peertube Die NGO AlgorithmWatch moniert, dass zentrale Forderungen aus Wissenschaft und Zivilgesellschaft unberücksichtigt geblieben seien. So fehle weiterhin ein verpflichtendes Transparenzregister für den KI-Einsatz in Behörden. Datenschutzbehörden hatten zudem gefordert, die Aufsicht über besonders sensible Hochrisiko-KI-Systeme ihnen statt der Bundesnetzagentur zu übertragen . Auch einen unabhängigen KI-Beirat hat das Parlament nicht gesetzlich verankert. Die stattdessen vorgesehene regelmäßige Beratung im Digitalausschuss gilt Kritikern als unzureichend. Der IT-Verband Bitkom warnt vor Problemen bei der Umsetzung. Da die Länder in den Vollzug eingebunden seien, drohten unterschiedliche Bewertungen identischer KI-Systeme und ein bürokratischer Flickenteppich. Die Bundesregierung verweist indes darauf, dass die Zuständigkeiten zwischen Bundesnetzagentur und Fachaufsichten wie der BaFin klar abgegrenzt werden könnten. Nach Angaben des Nationalen Normenkontrollrats entstehen dem Bund einmalige Kosten von rund vier Millionen Euro. Die laufenden jährlichen Ausgaben werden auf 15,9 Millionen Euro für den Bund und 33,1 Millionen Euro für die Länder geschätzt. Bei Verstößen gegen Mitwirkungs- und Auskunftspflichten drohen Bußgelder von bis zu 50.000 Euro. Weiterlesen nach der Anzeige ( mki )"
  },
  {
    "title": "Cohere veröffentlicht North Mini Code: 30B Open-Source-Modell für agentisches Coding",
    "summary": "Cohere hat mit North Mini Code ein neues Open-Source-Modell speziell für agentische Softwareentwicklung veröffentlicht. Das 30B-Parameter-Mixture-of-Experts-Modell (3B aktive Parameter) ist unter Apache-2.0-Lizenz auf HuggingFace verfügbar und übertrifft im Coding Index größere Modelle wie Nemotron 3 Super (120B) und Mistral Small 4 (119B). Das Modell wurde mit einer zweistufigen SFT-Pipeline und Reinforcement Learning mit verifizierbaren Rewards (RLVR) auf über 70.000 verifizierbaren Aufgaben aus rund 5.000 Repositories trainiert. Es ist für terminalbasierte Agenten-Workflows, komplexe Code-Generierung und Multi-Harness-Robustheit optimiert.",
    "source_url": "https://huggingface.co/blog/CohereLabs/introducing-north-mini-code",
    "tag": "tag-ki",
    "full_text": "Today, we are releasing North Mini Code, a 30B-parameter Mixture-of-Experts model with 3B active parameters with powerful agentic coding capabilities, available on Hugging Face under the Apache 2.0 license. North Mini Code is the first model in Cohere’s new family of models, and is specifically designed and trained for agentic software engineering tasks. Figure 1: North Mini Code’s performance in agentic coding tasks and complex code generation benchmarks, compared to leading open-source models of similar size. See here for the details of our benchmarking methodology. North Mini Code is optimized for complex software engineering workflows, terminal-based agentic tasks, and high-quality code generation. On Artificial Analysis’ Coding Index, North Mini Code achieves a score of 33.4, outperforming Qwen3.5 (35B-A3B), Gemma 4 (26B-A4B), Devstral Small 2 (24B Dense), and even substantially larger models such as Nemotron 3 Super (120B-A12B), Mistral Small 4 (119B-A6B), and Devstral 2 (123B). 1 It ranks among the strongest open-source coding models in its size class. Try North Mini Code in OpenCode Real-world code agents depend on model quality and robustness across agent harnesses. We trained North Mini Code using multiple scaffolds rather than optimizing for a single one. This approach enables North Mini Code to serve as a reliable foundation for coding agents such as OpenCode. Architecture Figure 2: North Mini Code is a Mixture-of-Experts Transformer decoder with interleaved sliding-window self-attention and full self-attention. North Mini Code is a decoder-only Transformer-based sparse Mixture-of-Experts model. It uses our efficient attention implementation, interleaved between sliding-window attention with RoPE and global attention with no positional embeddings, in a 3:1 ratio [ 1 ]. The feed-forward block is an MoE block with 128 experts, of which 8 are activated per token. Each expert block is an FFN block with SwiGLU activation. The router applies a sigmoid activation function to the logits before the top-k selection. We also use a single dense layer before the sparse layers. Post-Training for Coding Excellence Figure 3: The post-training pipeline is made up of two phases of supervised fine-tuning (SFT) and a phase of agentic reinforcement learning with verifiable rewards (RLVR) targeting software engineering and terminal tasks. We post-train North Mini Code using a two-stage cascaded supervised fine-tuning (SFT) followed by reinforcement learning with verifiable rewards (RLVR), focusing on agentic coding. Our first stage SFT data focuses on coding capabilities that are integrated within a wider mix for robustness and usability. The datamix includes programming, reasoning, and instruction following across a large variety of domains where the code datasets correspond to 70% of trainable tokens, 43% agentic tool-use data, and 27% single-turn competitive or scientific programming data. In the second stage SFT, we use a 4.5 billion token data mixture from only agentic and reasoning-driven samples, where code data forms 61% of trainable tokens. This mixture comprises our highest-quality data across coding and wider agentic tasks where tool calls and completions are verified as executable and correct. Our internal data pipeline heavily relies on containerised agentic coding environments. We maintain a disjoint subset of these environments for use in synthetic SFT data generation and RLVR. The majority are based on software engineering tasks from real-world repositories, while the rest are terminal-based agentic tasks sourced from open-source and internal datasets. In total, we used over 70k verifiable tasks across ~5k unique repositories. We deduplicate our environments against the repository sources from SWE-Bench [ 2 ] and SWE-Bench-Pro [ 3 ] to avoid source leakage during evaluation [ 4 ]. We used 64K and 128K context lengths for the first and second stages of SFT, respectively. This “long-to-longer” cascade approach (similar to [ 5 , 6 ]) enables bipartite training on valuable shorter data, establishing a robust performance baseline, followed by targeted long-context training only on high-quality verified samples. Without multi-stage training, the 20B non-code tokens during the initial training stage often dominated the 1.5B tokens of high-quality code data in later training, producing poorer performance and higher behavioral conflicts from data trends differing between stages. Anecdotally, training on a near-complete length distribution of samples produced shorter final trajectories during evaluation than training on a truncated distribution up to 64K only. Instead of optimising North Mini Code towards quantitative metrics during SFT, we adopted an approach strictly using SFT as priming for RLVR. The data mixture optimises sampling diversity and pass@K (for high K) in downstream stages. We use sample-level filtering to remove any pathologies such as invalid tool calls, erroneous whitespace generation, malformed special tokens, or hallucinated citations. Artifacts or hyperparameters producing undesirable RLVR behaviours (e.g., low entropy, invalid structured generations) were pruned via ablations. The final SFT model achieves 80.2% pass@10 on SWE-Bench Verified [ 2 ] and 55.1% pass@10 on Terminal-Bench v2 [ 7 ]. Robustness Across Harnesses Harness robustness improves model usability in realistic software development settings, where agents encounter diverse and unpredictable tooling environments. These environments differ not just in prompting but in fundamental tool-use modality, For instance, SWE-Agent [ 8 ] exposes a relatively rich agent-CLI interface with specialized commands ( bash , str_replace_editor and submit tools) and templated observations; mini-SWE-agent [ 9 ] strips this down to a single bash tool, with raw stdout from shell as the only feedback; and OpenCode [ 10 ] uses fine-grained individually typed tools ( edit , grep , todowrite and task etc) returning structured JSON responses. Figure 4: To power a variety of agentic coding harnesses, North Mini Code is exposed to a variety of coding harnesses during the second SFT stage. We address cross-harness generalization by introducing a small amount of additional benchmark harness data (6% of the SFT mix, compared to 50% of the chosen SWE-Agent harness) during the second SFT stage. Specifically, this data mix yields a 10% gain on the evaluation with OpenCode harness while maintaining performance with SWE-Agent on SWE-Bench Verified, demonstrating that cross-harness transfer can be cheaply acquired without degrading benchmark performance. Notably, North-Code-Mini achieves 61.0% pass@1 using mini-SWE-Agent, where the improvement emerged for free in the cross-task, cross-harness settings, suggesting that harnesses with overlapping tool capabilities share enough representational structure for positive transfer. We also observe minimal data conflict when training on hybrid harness data, indicating that skills required by different harnesses are usually complementary rather than contradictory. Similarly, the official Terminal-Bench uses its own Terminus 2 harness, where all the agent-CLI interactions are communicated via plain-text chat turns (instead of native tool calling). In order to prime our models on Terminus 2, we include a small amount of data (less than 20%) in a plain-text format in the data mixture, which has proved sufficient for the model to naturally generalise across. Interestingly, we also find that it’s crucial to introduce sufficient variations in the various harnesses (akin to data augmentation) in order to force the model to properly establish the link between instructions and behaviours rather than simply regurgitating a fixed template without understanding, and this is especially important when the harnesses appear similar to each other. Asynchronous RL for Agentic Coding Coding-agent rollouts are long and highly variable in length, with the slowest trajectories routinely an order of magnitude longer than the median. A synchronous RL loop would idle the trainer waiting for those trials to be generated for every batch, so we decouple sampling from learning: a trainer runs alongside a vLLM sidecar that serves rollouts continuously . Policy weights are exported into vLLM every few learner steps (K=4), so the sampler is at most slightly off-policy at any moment. The residual mismatch is then corrected at the loss level. To unblock the learner process from waiting on the longest rollouts while simultaneously avoiding a misbalance of data distribution across tasks, we used a windowed First-in-First-Out (FIFO) queue (trainer↔sampler) [ 11 ]: a small fraction at the head of the queue is consumed in completion order to drain stragglers, with the rest staying in input order. Empirically, this recovers most of the throughput of a completion-order scheme without measurably hindering training stability. We train using CISPO [ 12 ], a log-likelihood objective with token-level importance sampling correction. CISPO differs from PPO and GRPO in that the importance weight multiplies a log-likelihood rather than a probability ratio and enhances RLOO [ 13 ] with stronger regularization. We aggregate the loss at the token level rather than the prompt level, so the gradient signal scales with trajectory length and long agentic traces (where most of the credit-assignment signal lives) are not down-weighted relative to short ones. A single multi-environment RL train – We run a single multi-environment online RL training run spanning two task environments: Terminal-based tasks and software engineering tasks. Each training batch consists of 512 rollouts with a group size of 8 rollouts sampled per prompt. All rollouts share a global context window of 128K tokens. To account for differing task complexity, each task is assigned a distinct agentic-step budget. These per-task budgets were set based on pass@k filtering performed prior to RLVR, ensuring the budgets are appropriately calibrated to the difficulty of each task distribution. We observe that granting the model a turn budget substantially larger than necessary encourages unnecessary verbosity and hoppiness in its rollouts. For Terminal-based tasks, we configure the agent with a simple ReAct harness employing a single terminal-use tool based on Harbor's Tmux session implementation [ 14 ], whereas for SWE tasks, we employ the SWE-agent [ 8 ] harness. Both environments provide the agent with a pre-built Docker image encoding the environment state, a natural language user prompt, and a set of unit tests used for verification. We train on a combination of internal and open-source datasets, filtered to retain only problems with an acceptable pass@k rate, i.e., excluding trivially solved and completely unsolvable instances. We use binary rewards derived from the unit-test-based verifier. In addition, the model receives a reward of 0 for generating invalid tool calls or unparseable outputs, enabling a sharp drop in the rate of hallucinated or malformed tool calls within the first training steps. Figure 5: The multi-environment RL training run improves model performance on benchmarks like SWE-Bench Verified and Terminal-Bench v2. Learning curves are displayed on the left across the RLVR training process. Higher performance and robustness with online RL – RLVR training improved the performance of the final model from the SFT initialization by 7.9% (absolute) pass@1 in Terminal-Bench v2 and 3.0% (absolute) in SWE-Bench. We observe that joint training across both environments yields stronger results than training on each separately, and also generalizes better to out-of-distribution tasks. Beyond correctness scores, we observe significant improvements in agent robustness where the RLVR model produces shorter trajectories and fewer invalid or failing tool calls. The final model also exhibits less repetitive tool-call looping, reliably concluding its trajectory by submitting a solution or responding to the user. Internal Human Evaluation Benchmark Complementary to existing coding benchmarks, we also developed our own internal benchmark suite to measure model performance on out-of-distribution problems in pairwise evaluation with human annotators. In line with other benchmark setups, we evaluated the iterations of our models harnessed in OpenCode through Harbor. To understand model performance, we benchmark on four distinct functionalities: Code Explanation: Models are asked to explain particular technical aspects of a given code repository within a README file, or directly to the user. Code Editing: Models are tasked to implement a feature based on an existing code base. Data Visualization: Given data samples, models are tasked to create certain visualizations with a particular framework; no additional code is given. Implementation from Scratch: Given only design specifications and the packages to use, models are tasked to create a project from scratch, focused primarily on front-end design. Evaluators are provided with rubric-based scoring questions to help them assess individual response criteria and rate individual attempts first, before giving a final preference rating between the two model trajectories. 2 We share evaluation results of North Mini Code, comparing the SFT checkpoint with the final model release checkpoint. Figure 6: Pairwise preference results for human evaluation comparing the final North Mini Code checkpoint after RLVR against the SFT-only checkpoint across 85 samples. Our evaluations show that RLVR especially improves model performance on code editing tasks, resulting in an aggregate win rate of 66.1% across subsets for the final model against its SFT-only counterpart. Get Started North Mini Code models are available in OpenCode, Cohere API, and in HuggingFace with BF16 and FP8 (quantized) weights: bf16 , fp8 Extended Author List Code Agents Team and North Mini Code Group: Jay Alammar, Sophia Althammer, Dennis Aumiller, Leon Engländer, Yannis Flet-Berliac, Eden Gilbert, Sarra Habchi, Kylie He, Dhruti Joshi, Jozef Mokrý, David Mora, Josh Netto-Rosen, Deniz Qian, Lawrence Rodgers, Willem Röpke, Tom Sherborne, Ahmet Üstün, Minjie Xu Pre-training and Inference Team: Diana Abagyan, Sammie Bae, Björn Bebensee, Walter Beller-Morales, Sepideh Shaterian Bidgoli, Bas Büller, David Cairuz, Kris Cao, Roman Castagné, Giannis Chatziveroglou, Tim Chung, Felipe Cruz, Rishit Dholakia, Ali Edalati, Nikolas Gritsch, Kilian Haefeli, Prashant Kumar, Simon Lehnerer, Tony Liu, Alex McKinney, Ekagra Ranjan, Dev Shah, Zewen Shen, Sylvie Shi, Dwarak Talupuru, Komal Teru, Robin Vaaler, Bharat Venkitesh, Donglu Wang, Terrence Zhao, Leo Zhou, Conway Zhu Management and Leadership: Phil Blunsom, Nick Frosst, Aidan Gomez, Manoj Govindassamy, Nick Jakobi, Patrick Lewis, Acyr Locatelli, Joelle Pineau, Ivan Zhang Benchmarking Methodology Our core agentic capabilities are measured using SWE-Bench Verified, SWE-Bench Pro, Terminal-Bench v2, and Terminal-Bench Hard. North-Code-Mini was evaluated, using the Swe-Agent harness v1.1.0 for SWE-Bench, and a simple ReAct harness employing a single terminal-use tool based on Harbor’s Tmux session implementation for Terminal-Bench v2. For Terminal Bench Hard, we directly used Terminus-2, following the same methodology as the Artificial Analysis Intelligence Index to compare North Mini Code with the other models. We follow benchmarks’ official timeout and hardware resource limit settings wherever specified. We additionally track code generation capabilities in SciCode [ 15 ], which measures coding performance for scientific problems, and LiveCodeBench v6 [ 16 ], which requires strong algorithmic reasoning capabilities for coding performance outside of tool use. We run each benchmark with 3 different seeds and report the average benchmark performance, using temperature=1.0 and top_p=0.95. Competitor results – We used publicly reported scores for competitor models, either from original reports or the Artificial Analysis Intelligence Index, where available. Additionally, Gemma4’s scores for agentic coding tasks were reported by Qwen team \\[ 17 ]. For benchmark results that any public report is missing, denoted by (*) in Figure 1, we run them internally using the recommended model configuration. Citation @misc{cohere_north_code_mini, title = {Introducing {North Mini Code}: Cohere's First Model For Developers}, url = {cohere.com/blog/north-mini-code}, author = {{Team Cohere}}, month = {June}, year = {2026} } References [1] RoPE to NoPE and Back Again: A New Hybrid Attention Strategy [2] SWE-bench: Can Language Models Resolve Real-World GitHub Issues? [3] SWE-Bench Pro: Can AI Agents Solve Long-Horizon Software Engineering Tasks? [4] On Leakage of Code Generation Evaluation Datasets [5] Nemotron-Cascade: Scaling Cascaded Reinforcement Learning for General-Purpose Reasoning Models [6] Nemotron-Cascade 2: Post-Training LLMs with Cascade RL and Multi-Domain On-Policy Distillation [7] Terminal-Bench: A Benchmark for AI Agents in Terminal Environments [8] SWE-agent: Agent-Computer Interfaces Enable Automated Software Engineering [9] https://github.com/SWE-agent/mini-swe-agent [10] https://github.com/anomalyco/opencode [11] Forge: Scalable Agent RL Framework and Algorithm [12] MiniMax-M1: Scaling Test-Time Compute Efficiently with Lightning Attention [13] Back to Basics: Revisiting REINFORCE-Style Optimization for Learning from Human Feedback in LLMs [14] Harbor: A Framework for Evaluating and Optimizing Agents and Models in Container Environments [15] SciCode: A Research Coding Benchmark Curated by Scientists [16] LiveCodeBench: Holistic and Contamination Free Evaluation of Large Language Models for Code [17] Qwen3.6-35B-A3B: Agentic Coding Power, Now Open to All Footnotes 1. AAII Coding Index includes Terminal Bench Hard as an agentic coding task and SciCode as code generation benchmark for scientific problems. ↩ 2. Both individual ratings and preferences are assessed on a five-point Likert scale. ↩ Models mentioned in this article 2 More from this author Build Small Hackathon With Cohere Models 5 June 4, 2026 Introducing Cohere-transcribe: state-of-the-art speech recognition 46 March 26, 2026 Community Congrats !!! </p>\\n\",\"updatedAt\":\"2026-06-09T16:21:03.373Z\",\"author\":{\"_id\":\"6168218a4ed0b975c18f82a8\",\"avatarUrl\":\"https://cdn-avatars.huggingface.co/v1/production/uploads/6168218a4ed0b975c18f82a8/Qvx6BqnR1S_Q8Ajz_eKRX.png\",\"fullname\":\"NIONGOLO Chrys Fé-Marty\",\"name\":\"Svngoku\",\"type\":\"user\",\"isPro\":true,\"isHf\":false,\"isHfAdmin\":false,\"isMod\":false,\"followerCount\":85,\"isUserFollowing\":false,\"primaryOrg\":{\"avatarUrl\":\"https://cdn-avatars.huggingface.co/v1/production/uploads/60d2dc1007da9c17c72708f8/XejxNYtiveJxFYcILIEBh.png\",\"fullname\":\"Build Small Hackathon\",\"name\":\"build-small-hackathon\",\"type\":\"org\",\"isHf\":false,\"plan\":\"team\"}}},\"numEdits\":0,\"identifiedLanguage\":{\"language\":\"de\",\"probability\":0.31143227219581604},\"editors\":[\"Svngoku\"],\"editorAvatarUrls\":[\"https://cdn-avatars.huggingface.co/v1/production/uploads/6168218a4ed0b975c18f82a8/Qvx6BqnR1S_Q8Ajz_eKRX.png\"],\"reactions\":[{\"reaction\":\"👍\",\"users\":[\"arbv\",\"SarraHab\",\"XGeorgeCostanzaX\"],\"count\":3}],\"isReport\":false}},{\"id\":\"6a2bbec543b64b323c0c294e\",\"author\":{\"_id\":\"636a9afec0be1b0cfbe4b3a6\",\"avatarUrl\":\"/avatars/5a37574d1f91add33b7eb655c1d8f902.svg\",\"fullname\":\"LEI WANG\",\"name\":\"yiakwy-xpu-team\",\"type\":\"user\",\"isPro\":false,\"isHf\":false,\"isHfAdmin\":false,\"isMod\":false,\"followerCount\":17,\"isUserFollowing\":false},\"createdAt\":\"2026-06-12T08:09:41.000Z\",\"type\":\"comment\",\"data\":{\"edited\":false,\"hidden\":false,\"latest\":{\"raw\":\"@coherecode Thank you for sharing! In our internal test DSV4 is really good as baseline , so it will be good that we could sft/RL over DSV4 (we have ran though whole pipeline in week 0 support, and it will take 2 days to finish trainining over 10k+ tool calls/coding corpus) ?\\n\\nI wonder if dsv4+ coding corpus can generate new record in coding (for agents).\\n\\n\",\"html\":\"<p><span class=\\\"SVELTE_PARTIAL_HYDRATER contents\\\" data-target=\\\"UserMention\\\" data-props=\\\"{\"user\":\"coherecode\"}\\\"><span class=\\\"inline-block\\\"><span class=\\\"contents\\\"><a href=\\\"/coherecode\\\">@<span class=\\\"underline\\\">coherecode</span></a></span> </span></span> Thank you for sharing! In our internal test DSV4 is really good as baseline , so it will be good that we could sft/RL over DSV4 (we have ran though whole pipeline in week 0 support, and it will take 2 days to finish trainining over 10k+ tool calls/coding corpus) ?</p>\\n<p>I wonder if dsv4+ coding corpus can generate new record in coding (for agents).</p>\\n\",\"updatedAt\":\"2026-06-12T08:09:41.400Z\",\"author\":{\"_id\":\"636a9afec0be1b0cfbe4b3a6\",\"avatarUrl\":\"/avatars/5a37574d1f91add33b7eb655c1d8f902.svg\",\"fullname\":\"LEI WANG\",\"name\":\"yiakwy-xpu-team\",\"type\":\"user\",\"isPro\":false,\"isHf\":false,\"isHfAdmin\":false,\"isMod\":false,\"followerCount\":17,\"isUserFollowing\":false}},\"numEdits\":0,\"identifiedLanguage\":{\"language\":\"en\",\"probability\":0.9020848274230957},\"editors\":[\"yiakwy-xpu-team\"],\"editorAvatarUrls\":[\"/avatars/5a37574d1f91add33b7eb655c1d8f902.svg\"],\"reactions\":[{\"reaction\":\"🔥\",\"users\":[\"Svngoku\"],\"count\":1}],\"isReport\":false}}],\"status\":\"open\",\"isReport\":false,\"pinned\":false,\"locked\":false,\"collection\":\"community_blogs\"},\"contextAuthors\":[\"coherecode\"],\"primaryEmailConfirmed\":false,\"discussionRole\":0,\"acceptLanguages\":[\"*\"],\"withThread\":true,\"cardDisplay\":false,\"repoDiscussionsLocked\":false,\"hideComments\":true}\"> Svngoku 4 days ago Congrats !!! 👍 3 3 + Reply yiakwy-xpu-team about 21 hours ago @ coherecode Thank you for sharing! In our internal test DSV4 is really good as baseline , so it will be good that we could sft/RL over DSV4 (we have ran though whole pipeline in week 0 support, and it will take 2 days to finish trainining over 10k+ tool calls/coding corpus) ? I wonder if dsv4+ coding corpus can generate new record in coding (for agents). See translation 🔥 1 1 + Reply Edit Preview Upload images, audio, and videos by dragging in the text input, pasting, or clicking here . Tap or paste here to upload images Comment · Sign up or log in to comment"
  },
  {
    "title": "Kimi K2.7 Code: Moonshot AI veröffentlicht stärkstes Coding-Modell mit 256K-Kontext",
    "summary": "Moonshot AI hat Kimi K2.7 Code veröffentlicht, das nach eigenen Angaben leistungsfähigste Coding-Modell des Unternehmens. Es bietet einen 256K-Token-Kontext, Durchbrüche bei langfristigen Coding-Aufgaben über diverse Programmiersprachen (Rust, Go, Python) und Szenarien (Frontend, DevOps, Performance-Optimierung) hinweg. Das Modell unterstützt multimodale Tool-Nutzung, kann Videos analysieren und komplexe mehrschrittige Reasoning-Aufgaben lösen. Die API ist OpenAI-kompatibel und über platform.kimi.ai verfügbar – mit einer Einführungsaktion.",
    "source_url": "https://platform.kimi.ai/docs/guide/kimi-k2-7-code-quickstart",
    "tag": "tag-ki",
    "full_text": "Kimi K2.7 Code, our most capable coding model to date. It follows instructions more reliably in long contexts, completes coding tasks with higher success rates. ​ Long-horizon coding capability breakthrough K2.7 Code has achieved a breakthrough in long-horizon coding tasks, demonstrating more reliable generalization across diverse programming languages (such as Rust, Go, and Python) and task scenarios (including frontend development, DevOps, and performance optimization). ​ Ultra-Long Context Support kimi-k2.7-code , kimi-k2.6 , kimi-k2.5 models all provide a 256K context window. ​ Long-Thinking Capabilities Kimi K2.7 Code still has strong reasoning capabilities, supporting multi-step tool invocation and reasoning, excelling at solving complex problems, such as complex logical reasoning, mathematical problems, and code writing. Kimi K2.7 Code does not support non-thinking mode. ​ Example Usage Here is a complete usage example to help you quickly get started with the Kimi K2.7 Code model. ​ Install the OpenAI SDK Kimi API is fully compatible with OpenAI’s API format. You can install the OpenAI SDK as follows: pip install -- upgrade 'openai>=1.0' ​ Verify the Installation python - c 'import openai; print(\"version =\",openai.__version__)' # The output may be version = 1.10.0, indicating the OpenAI SDK was installed successfully and your Python environment is using OpenAI SDK v1.10.0. ​ Quick Start Try it now : Test model performance in your business scenarios through interactive operations in the Dev Workbench Apply for API Key : Test via API call immediately ​ Multimodal Tool Capability Example Kimi K2.7 Code model combines multiple capabilities. The following example demonstrates K2.7 Code’s visual understanding + tool calling capabilities. First, download this sample video to your local machine, such as ~/Download/test_video.mp4 Then run the following code: import base64 import json import os import subprocess import tempfile from pathlib import Path from openai import OpenAI tools = [{ \"type\" : \"function\" , \"function\" : { \"name\" : \"watch_video_clip\" , \"description\" : \"Watch a video file or a sub-clip of it. If start_time and end_time are not provided, the entire video will be returned.\" , \"parameters\" : { \"type\" : \"object\" , \"properties\" : { \"path\" : { \"type\" : \"string\" , \"description\" : \"The path to the video file to watch\" }, \"start_time\" : { \"type\" : \"number\" , \"description\" : \"The start time of the clip in seconds (optional, defaults to 0)\" }, \"end_time\" : { \"type\" : \"number\" , \"description\" : \"The end time of the clip in seconds (optional, defaults to end of video)\" } }, \"required\" : [ \"path\" ] } } }] def watch_video_clip ( path : str , start_time : float | None = None , end_time : float | None = None ) -> list[ dict ]: \"\"\" Watch a video file or a sub-clip of it. Args: path: The path to the video file to watch start_time: The start time in seconds (optional, defaults to 0) end_time: The end time in seconds (optional, defaults to end of video) Returns: A list of content blocks in MultiModal Tool API format \"\"\" video_path = Path(path) if not video_path.exists(): raise FileNotFoundError ( f \"Video file not found: { path } \" ) # Get video duration if needed if start_time is None and end_time is None : # Return entire video with open (path, \"rb\" ) as f: video_base64 = base64.b64encode(f.read()).decode( \"utf-8\" ) return [ { \"type\" : \"video_url\" , \"video_url\" : { \"url\" : f \"data:video/mp4;base64, { video_base64 } \" }}, { \"type\" : \"text\" , \"text\" : f \"Full video: { video_path.name } \" } ] # Get video duration for defaults probe = subprocess.run( [ \"ffprobe\" , \"-v\" , \"quiet\" , \"-print_format\" , \"json\" , \"-show_format\" , path], capture_output = True , text = True ) duration = float (json.loads(probe.stdout)[ \"format\" ][ \"duration\" ]) start_time = start_time or 0 end_time = end_time or duration clip_duration = end_time - start_time # Extract clip with tempfile.NamedTemporaryFile( suffix = \".mp4\" , delete = False ) as tmp: tmp_path = tmp.name try : subprocess.run([ \"ffmpeg\" , \"-y\" , \"-ss\" , str (start_time), \"-i\" , path, \"-t\" , str (clip_duration), \"-c:v\" , \"libx264\" , \"-c:a\" , \"aac\" , \"-preset\" , \"fast\" , \"-crf\" , \"23\" , \"-movflags\" , \"+faststart\" , \"-loglevel\" , \"error\" , tmp_path ], check = True ) with open (tmp_path, \"rb\" ) as f: video_base64 = base64.b64encode(f.read()).decode( \"utf-8\" ) return [ { \"type\" : \"video_url\" , \"video_url\" : { \"url\" : f \"data:video/mp4;base64, { video_base64 } \" }}, { \"type\" : \"text\" , \"text\" : f \"Clip from { video_path.name } : { start_time } s - { end_time } s\" } ] finally : if os.path.exists(tmp_path): os.unlink(tmp_path) client = OpenAI( api_key = os.environ.get( \"MOONSHOT_API_KEY\" ), base_url = \"https://api.moonshot.ai/v1\" ) def agent_loop ( user_message : str ): \"\"\"Simple agent loop with multimodal tool support.\"\"\" messages = [ { \"role\" : \"system\" , \"content\" : \"You are a video analysis assistant. Use watch_video_clip to examine specific portions of videos.\" }, { \"role\" : \"user\" , \"content\" : user_message} ] while True : response = client.chat.completions.create( model = \"kimi-k2.7-code\" , messages = messages, tools = tools, tool_choice = \"auto\" ) message = response.choices[ 0 ].message messages.append(message.model_dump()) # No tool calls = done if not message.tool_calls: return message.content # Execute tool calls for tool_call in message.tool_calls: if tool_call.function.name == \"watch_video_clip\" : args = json.loads(tool_call.function.arguments) result = watch_video_clip( path = args[ \"path\" ], start_time = args.get( \"start_time\" ), end_time = args.get( \"end_time\" ) ) # Multimodal tool result messages.append({ \"role\" : \"tool\" , \"tool_call_id\" : tool_call.id, \"content\" : result }) # Usage answer = agent_loop( \"Analyze what happens between seconds 8-13 in ~/Download/test_video.mp4\" ) print (answer) ​ Best Practices ​ Supported Formats Images are supported in formats: png, jpeg, webp, gif. Videos are supported in formats: mp4, mpeg, mov, avi, x-flv, mpg, webm, wmv, 3gpp. ​ Token Calculation and Billing Image and video token usage is dynamically calculated. You can use the token estimation API to check the expected token consumption for a request containing images or video before processing. Generally, the higher the resolution of an image, the more tokens it will consume. For videos, the number of tokens depends on the number of keyframes and their resolution—the more keyframes and the higher their resolution, the greater the token consumption. The Vision model uses the same billing method as the moonshot-v1 model series, with charges based on the total number of tokens processed. For more information, see: For token pricing details, refer to Model Pricing . ​ Recommended Resolution We recommend that image resolution should not exceed 4k (4096×2160), and video resolution should not exceed 2k (2048×1080). Higher resolutions will only increase processing time and will not improve the model’s understanding. ​ Upload File or Base64? Due to the limitation on the overall size of the request body, for very large videos you must use the file upload method to utilize vision capabilities.For images or videos that will be referenced multiple times, it is recommended to use the file upload method. Regarding file upload limitations, please refer to the File Upload documentation . Image quantity limit: The Vision model has no limit on the number of images, but ensure that the request body size does not exceed 100M URL-formatted images: Not supported, currently only supports base64-encoded image content ​ Parameters Differences in Request Body Parameters are listed in chat . However, behaviour of some parameters may be different in k2.7-code/k2.6/k2.5 models. We recommend using the default values instead of manually configuring these parameters. Differences are listed below. Field Required Description Type Values max_tokens optional The maximum number of tokens to generate for the chat completion. int Default to be 32k aka 32768 thinking optional New! This parameter controls if the thinking is enabled for this request object Default to be {\"type\": \"enabled\"} . Kimi K2.7 Code model will throw an error if the thinking mode is disabled. temperature optional The sampling temperature to use float Kimi K2.7 Code model will use a fixed value 1.0. Any other value will result in an error top_p optional A sampling method float Kimi K2.7 Code model will use a fixed value 0.95. Any other value will result in an error n optional The number of results to generate for each input message int Kimi K2.7 Code model will use a fixed value 1. Any other value will result in an error presence_penalty optional Penalizing new tokens based on whether they appear in the text float Kimi K2.7 Code model will use a fixed value 0.0. Any other value will result in an error frequency_penalty optional Penalizing new tokens based on their existing frequency in the text float Kimi K2.7 Code model will use a fixed value 0.0. Any other value will result in an error ​ Tool Use Compatibility When using tools, please note the following constraints to ensure model performance: tool_choice can only be set to “auto” or “none” (default is “auto”) to avoid conflicts between reasoning content and the specified tool_choice. Any other value will result in an error; During multi-step tool calling, you must keep the reasoning_content from the assistant message in the current turn’s tool call within the context, otherwise an error will be thrown; ​ Model Pricing For token pricing details, refer to Model Pricing . ​ Learn More For the benchmark testing with Kimi K2.7 Code, please refer to this benchmark best practice For the most detailed API usage example of Kimi K2.7 Code, see: How to Use Kimi Vision Model See How to Use Kimi K2 in Claude Code, Roo Code, and Cline Learn how to configure and use the Thinking Model For all model pricing see here , Billing & Rate Limit details , and Web Search Pricing Was this page helpful? Yes No Quickstart with Kimi API Kimi K2.6 Multi-modal Model ⌘ I x github linkedin"
  }
]