Hermes Agent
193 lines
10 KiB
Markdown
193 lines
10 KiB
Markdown
# KI/LLM-News für Basti's Morning Briefing — 21. Juni 2026
|
||
|
||
---
|
||
|
||
## Story 1: GLM-5.2 — Chinas neues Open-Source-Flaggschiff mit 1M-Kontext
|
||
|
||
- **headline:** GLM-5.2: Z.AI veröffentlicht 753B-Parameter-Open-Source-Modell mit 1M-Token-Kontext unter MIT-Lizenz
|
||
- **tag:** tag-ki
|
||
- **teaser:** Z.AI (Zhipu AI) hat am 16. Juni GLM-5.2 veröffentlicht — ein 753-Milliarden-Parameter-MoE-Modell mit stabilem 1M-Token-Kontextfenster. Das Modell erreicht auf Long-Horizon-Coding-Benchmarks nahezu Opus-4.8-Niveau (FrontierSWE: 74.4 vs. 75.1) und schlägt GPT-5.5. Es steht unter MIT-Lizenz auf HuggingFace, ist via Cloudflare Workers AI, FriendliAI und Novita nutzbar und kostet nur $1.40/$4.40 pro Million Tokens.
|
||
- **source_url:** https://z.ai/blog/glm-5.2
|
||
- **source_name:** Z.AI Blog
|
||
- **full_text:** PAYWALL (vollständiger Text via web_extract scrapbar, siehe unten)
|
||
- **category:** Open-Source
|
||
|
||
---
|
||
|
||
## Story 2: Claude Fable 5 — US-Regierung zwingt Anthropic zur Abschaltung
|
||
|
||
- **headline:** US-Exportdirektive zwingt Anthropic zur globalen Abschaltung von Claude Fable 5 und Mythos 5
|
||
- **tag:** tag-ki
|
||
- **teaser:** Nur 72 Stunden nach dem Launch von Claude Fable 5 am 9. Juni ordnete das US-Handelsministerium am 12. Juni per Exportkontroll-Direktive die vollständige Sperrung für alle ausländischen Nutzer an — faktisch eine globale Abschaltung. Es ist das erste Mal, dass die US-Regierung ein bereits live geschaltetes Frontier-KI-Modell per Exportkontrolle vom Netz nimmt. Der Präzedenzfall wirft fundamentale Fragen zur Abhängigkeit von einzelnen Cloud-API-Anbietern auf.
|
||
- **source_url:** https://www.anthropic.com/news/claude-fable-5-mythos-5
|
||
- **source_name:** Anthropic News
|
||
- **full_text:** PAYWALL (vollständiger Text via web_extract scrapbar, siehe unten)
|
||
- **category:** Cloud-API
|
||
|
||
---
|
||
|
||
## Story 3: Kimi K2.7-Code — Moonshot AIs 1T-Parameter Coding-Modell
|
||
|
||
- **headline:** Kimi K2.7-Code: Moonshot AI veröffentlicht 1-Billion-Parameter Open-Weight Coding-Modell
|
||
- **tag:** tag-ki
|
||
- **teaser:** Moonshot AI hat am 12. Juni Kimi K2.7-Code released — ein 1T-Parameter-MoE-Modell (32B aktiv) mit Fokus auf Long-Horizon-Software-Engineering. Es erreicht +21,8% auf dem Kimi Code Bench v2 gegenüber K2.6 und schlägt Claude Opus 4.8 auf MCP Mark Verified (81.1 vs. 76.4). Die Open Weights (Modified MIT) liegen auf HuggingFace, API-Pricing liegt bei $0.95/$4.00 pro Million Tokens — etwa 80% günstiger als Opus 4.8.
|
||
- **source_url:** https://www.marktechpost.com/2026/06/12/moonshot-ai-releases-kimi-k2-7-code-a-coding-model-reporting-21-8-on-kimi-code-bench-v2-over-k2-6
|
||
- **source_name:** MarkTechPost
|
||
- **full_text:** PAYWALL (vollständiger Text via web_extract scrapbar, siehe unten)
|
||
- **category:** Open-Source
|
||
|
||
---
|
||
|
||
## Story 4: DiffusionGemma — Googles 4x schnelleres Text-Diffusion-Modell
|
||
|
||
- **headline:** DiffusionGemma: Google veröffentlicht experimentelles Open-Source-Modell mit 4x schnellerer Textgenerierung
|
||
- **tag:** tag-ki
|
||
- **teaser:** Google DeepMind hat am 10. Juni DiffusionGemma veröffentlicht — ein 26B-MoE-Modell (3.8B aktiv), das Text nicht sequenziell, sondern per Diffusion in 256-Token-Blöcken parallel generiert. Auf einer NVIDIA H100 erreicht es über 1000 Tokens/Sekunde, auf einer RTX 5090 über 700 t/s. Es läuft in ~18 GB VRAM (quantisiert) und eignet sich besonders für Code-Infilling und nicht-lineare Textstrukturen. Apache-2.0-Lizenz, HuggingFace, vLLM-Support.
|
||
- **source_url:** https://blog.google/innovation-and-ai/technology/developers-tools/diffusion-gemma-faster-text-generation/
|
||
- **source_name:** Google Blog
|
||
- **full_text:** PAYWALL (vollständiger Text via web_extract scrapbar, siehe unten)
|
||
- **category:** Open-Source
|
||
|
||
---
|
||
|
||
## Story 5: "Running local models is good now" — Lokale LLMs erreichen Praxistauglichkeit
|
||
|
||
- **headline:** "Running local models is good now": Lokale LLMs erreichen 75% der Frontier-Modelle — ein Praxisbericht
|
||
- **tag:** tag-ki
|
||
- **teaser:** Vicki Boykis' Blogpost vom 15. Juni wurde mit über 1.550 Punkten auf Hacker News zur meistdiskutierten Story der Woche. Sie beschreibt, wie Modelle wie Gemma 4 12B und Qwen 3.6 auf einem M2 Mac mit 64 GB RAM agentisches Coding mit ~75% der Genauigkeit von Cloud-Frontier-Modellen ermöglichen — Tasks, die vor 6 Monaten noch unmöglich waren. Der Artikel enthält konkrete Docker-Compose-Setups für den Open-Source-Agenten-Harness "Pi" mit lokalem LM Studio.
|
||
- **source_url:** https://vickiboykis.com/2026/06/15/running-local-models-is-good-now/
|
||
- **source_name:** Vicki Boykis (Blog)
|
||
- **full_text:** PAYWALL (vollständiger Text via web_extract scrapbar, siehe unten)
|
||
- **category:** Tools
|
||
|
||
---
|
||
|
||
## Vollständige Artikeltexte (via web_extract gescrapt)
|
||
|
||
### GLM-5.2 (z.ai/blog/glm-5.2)
|
||
```
|
||
GLM-5.2: Built for Long-Horizon Tasks
|
||
Source: Z.ai Blog | Date: 2026-06-16
|
||
|
||
Key Highlights:
|
||
- Solid 1M-token context – stable, engineering-usable long context
|
||
- Advanced Coding with Flexible Effort – multiple thinking effort levels (High, Max)
|
||
- Improved Architecture (IndexShare) – reduces per-token FLOPs by 2.9× at 1M context
|
||
- MTP layer improvements – acceptance length increased by 20%
|
||
- Pure Open – MIT license, no regional limits
|
||
- Top open-source model on long-horizon coding benchmarks
|
||
|
||
Performance Benchmarks (Long-Horizon Coding):
|
||
| Benchmark | GLM-5.2 | GLM-5.1 | Opus 4.8 | GPT-5.5 |
|
||
| FrontierSWE | 74.4 | 30.5 | 75.1 | 72.6 |
|
||
| PostTrainBench | 34.3 | 20.1 | 37.2 | 28.4 |
|
||
| SWE-Marathon | 13.0 | 1.0 | 26.0 | 12.0 |
|
||
|
||
Standard Coding:
|
||
| Terminal-Bench 2.1 | 81.0 | 63.5 | 85.0 | - |
|
||
| SWE-bench Pro | 62.1 | 58.4 | 69.2 | - |
|
||
|
||
Reasoning:
|
||
| AIME 2026 | 99.2 | 95.3 | 95.7 | 98.3 |
|
||
| GPQA-Diamond | 91.2 | 86.2 | 93.6 | 93.6 |
|
||
|
||
Architecture: 753B total parameters, MoE. IndexShare reuses indexer across every 4 sparse attention layers. MTP with KVShare increases acceptance length from 4.56 to 5.47. Anti-hack system for coding RL prevents reward hacking during training.
|
||
|
||
Pricing: $1.40/M input, $4.40/M output. Available via FriendliAI, Novita, ZAI, Cloudflare Workers AI.
|
||
License: MIT (open-weight, commercial use allowed)
|
||
```
|
||
|
||
### Claude Fable 5 Suspension (anthropic.com/news/claude-fable-5-mythos-5)
|
||
```
|
||
Claude Fable 5 and Claude Mythos 5 – Anthropic
|
||
Date: June 9, 2026 | Update (Jun 12): Access suspended
|
||
|
||
Overview:
|
||
- Claude Fable 5: Mythos-class model made safe for general use. SOTA on nearly all benchmarks.
|
||
- Claude Mythos 5: Same model with safeguards lifted (cyberdefenders via Project Glasswing).
|
||
- Pricing: $10/M input, $50/M output (less than half Mythos Preview).
|
||
|
||
Capabilities:
|
||
- Stripe: Compressed months of engineering into days – codebase-wide migration in 50M-line Ruby codebase in one day.
|
||
- FrontierCode: Highest score among frontier models at medium effort.
|
||
- Hebbia Finance Benchmark: Highest score of any model.
|
||
- Vision: New SOTA. Rebuilds web app source code from screenshots. Beat Pokémon FireRed with minimal vision-only harness.
|
||
- Drug Design (Mythos 5): Accelerated aspects ~10×. 9 of 14 protein targets yielded strong candidates.
|
||
- Novel Hypotheses: First model to consistently produce novel, compelling scientific hypotheses. Scientists preferred Mythos hypotheses ~80% of the time.
|
||
|
||
Safeguards: New classifiers detect potential misuse → fall back to Opus 4.8. >95% of sessions involve no fallback. Covers cybersecurity, biology/chemistry, distillation.
|
||
|
||
Suspension: On June 12, 2026, US Commerce Department issued export control directive ordering Anthropic to suspend all access to Fable 5 and Mythos 5 for foreign nationals → effectively global shutdown. First time US government used export controls to force a frontier AI model offline post-launch.
|
||
```
|
||
|
||
### Kimi K2.7-Code (MarkTechPost)
|
||
```
|
||
Moonshot AI Releases Kimi K2.7-Code – June 12, 2026
|
||
Model: 1T-parameter MoE (32B active), 384 experts (8+1 per token), 256K context
|
||
License: Modified MIT (open weights on Hugging Face)
|
||
|
||
Benchmarks:
|
||
| Benchmark | K2.6 | K2.7-Code | GPT-5.5 | Opus 4.8 | Improvement |
|
||
| Kimi Code Bench v2 | 50.9 | 62.0 | 69.0 | 67.4 | +21.8% |
|
||
| Program Bench | 48.3 | 53.6 | 69.1 | 63.8 | +11.0% |
|
||
| MLS Bench Lite | 26.7 | 35.1 | 35.5 | 42.8 | +31.5% |
|
||
| MCP Mark Verified | 72.8 | 81.1 | 92.9 | 76.4 | +11.4% |
|
||
|
||
Pricing: $0.95/M input, $4.00/M output (cached: $0.19). ~80% cheaper than Opus 4.8.
|
||
Key: 30% lower reasoning-token usage than K2.6. Mandatory thinking mode. Fixed sampling params.
|
||
```
|
||
|
||
### DiffusionGemma (Google Blog)
|
||
```
|
||
DiffusionGemma: 4x Faster Text Generation – June 10, 2026
|
||
License: Apache 2.0
|
||
|
||
Architecture: 26B MoE (3.8B active). Text diffusion instead of autoregressive – generates 256-token blocks in parallel, iteratively refined.
|
||
|
||
Performance:
|
||
- NVIDIA H100: 1000+ tokens/sec
|
||
- NVIDIA RTX 5090: 700+ tokens/sec
|
||
- VRAM: ~18GB (quantized)
|
||
|
||
Key innovation: Bi-directional attention – all 256 tokens attend to each other simultaneously. Enables non-linear text structures (code infilling, math graphs, amino acid sequences).
|
||
|
||
Trade-offs: Lower output quality than standard Gemma 4. Best for low-concurrency local inference. Experimental status.
|
||
|
||
Supported: MLX, vLLM, Hugging Face Transformers, Unsloth, NVIDIA NeMo, llama.cpp (soon). Optimized for RTX 5090/4090, Hopper/Blackwell (NVFP4).
|
||
```
|
||
|
||
### "Running local models is good now" (Vicki Boykis)
|
||
```
|
||
Running Local Models Is Good Now – June 15, 2026
|
||
Author: Vicki Boykis | Hardware: 2022 M2 Mac, 64 GB RAM
|
||
|
||
Key insight: With Gemma 4 family, agentic coding locally now works at ~75% accuracy/speed of frontier models.
|
||
|
||
Tasks accomplished locally:
|
||
- Refactored Python notebook into 5-6 module repo
|
||
- Linted code for PEP 585 type hints
|
||
- Proofread blog posts, wrote unit tests
|
||
- Bootstrapped two-tower recommendation model repo
|
||
- Built app surfacing trending Arxiv topics
|
||
|
||
Setup: Pi (harness) + LM Studio (inference) + gemma-4-12b-qat (model). All Pi sessions run in Docker with restricted permissions.
|
||
|
||
Remaining issues: Slow inference (KV cache up to 64 GB RAM), small context windows, prompt template mismatches on early releases.
|
||
|
||
HN: 1554 points, 596 comments – #1 on Hacker News.
|
||
```
|
||
|
||
---
|
||
|
||
## Quellenübersicht
|
||
|
||
| # | Story | Quelle | Datum | Kategorie |
|
||
|---|-------|--------|-------|-----------|
|
||
| 1 | GLM-5.2 Release | z.ai/blog/glm-5.2 | 16.06.2026 | Open-Source |
|
||
| 2 | Claude Fable 5 Suspension | anthropic.com/news | 12.06.2026 | Cloud-API |
|
||
| 3 | Kimi K2.7-Code | marktechpost.com | 12.06.2026 | Open-Source |
|
||
| 4 | DiffusionGemma | blog.google | 10.06.2026 | Open-Source |
|
||
| 5 | Local Models Are Good Now | vickiboykis.com | 15.06.2026 | Tools |
|
||
|
||
*Recherche durchgeführt am 21. Juni 2026. Keine neuen Major-Releases am 19.-21. Juni gefunden; die obigen Stories sind die relevantesten der letzten 5-10 Tage.*
|