Fresh stories
Meituan releases LongCat-2.5 multimodal agent API with 1.6T parameters
Meituan released LongCat-2.5 through an API, with 1.6T total parameters and about 48B active. Launch materials describe agentic reinforcement learning, KV-cache reuse and interactive coding demos.
DeepSeek V4.1 Flash 1017 scores 5.81% on ARC-AGI-3 with provider adapter
ARC Prize verified DeepSeek V4.1 Flash 1017 at 5.81% on ARC-AGI-3 with the provider adapter, versus 0.2% with the standard harness. Its semi-private score more than doubled its predecessor’s.
FineDNA releases 514B training tokens and 300,000 biology-model tasks
FineDNA released 514B mid-training tokens, 200,000 SFT conversations and 100,000 RL tasks. Its authors report about 67% variant-classification accuracy with code-grounded explanations, versus below 60% for self-distillation.
Meituan releases LongCat-2.5 multimodal agent API with 1.6T parameters
Meituan released LongCat-2.5 through an API, with 1.6T total parameters and about 48B active. Launch materials describe agentic reinforcement learning, KV-cache reuse and interactive coding demos.
Claude API adds server-managed message threads
Anthropic’s public beta lets clients continue cached conversations by sending a previous message ID. Anthropic reports 35% lower median time to first token on Claude Code requests with 500K–1M tokens, at standard token rates.
Unsloth reports UD3.1 beats GSQ-RCO in a revised 35.1B-token quantization test
Unsloth says its UD3.1 GGUFs beat GSQ-RCO in a 35.1B-token rerun using longer contexts and revised grading. It says the competing team’s benchmark methods remain undisclosed.
Zyphra adds tree-based routing to Hierarchical Attention for sparse retrieval
Zyphra Hierarchical Attention adds learned tree routing to pretrained models for sparse retrieval. It retains the full KV cache; the large speedup figures are roofline projections, not measured serving gains.
DeepSeek V4.1 Flash 1017 scores 5.81% on ARC-AGI-3 with provider adapter
OpenAI says Pando cuts monorepo Git push latency by 80%
MIT researchers use Astra to raise Qwen MATH-500 accuracy from 7.9% to 48.1%
FineDNA releases 514B training tokens and 300,000 biology-model tasks
Top storiesthis week
SemiAnalysis releases GDPx benchmark; Opus 5.5 leads at 12%
SemiAnalysis released GDPx to evaluate agents on economically valuable tasks, with Opus 5.5 leading at 12.0%. The publisher says safety classifiers materially affect some scores, complicating model comparisons.
DeepSeek releases dsh-superagent for GitHub repository tasks
DeepSeek's dsh-superagent coordinates isolated workers through GitHub issues and pull requests. A reported internal test cut p95 task completion time from 29.5 to 15.7 minutes versus a session using subagents.
T3 Code usage comparisons draw criticism over unmatched tasks
Theo says T3 Code's comparisons aggregate real-world usage rather than matched tasks. Critics cite prompts, effort settings, custom tools and cache handling as factors that could affect the results.
Karpathy builds HackerTube from 43 million Hacker News comments
Karpathy says he built HackerTube in about two hours with Claude Code. Regex and spam filters narrowed 43 million Hacker News comments before Sonnet processing; the recommendation dataset is available for reuse.
EdgeRock-4B fixes its attention KV cache at 72 MiB
Rock Matrix released EdgeRock-4B with a 72 MiB attention KV cache built around local attention and memory anchors. Weights, code, and quantized versions are available, along with a local laptop demonstration.