[AINews] not much happened today

What it claims

This AINews issue (Latent Space, 16-19 October 2026) is a news digest, not a primary source. Its most actionable item is DeepSeek V4.1 Flash 1017. That release keeps the prior price of $0.30/$1.20 per million input/output tokens at peak hours, half that off-peak, and $0.006 per million for cache hits. Artificial Analysis (AA) ranks it tied #2 among open-weight models at $0.26 per task. Together says DeepSeek trained the update with self-feedback RL on building complete projects from scratch. The issue also reports MiniMax M3.1 Flash pricing and weights timing, the vLLM v0.32.0 feature set, the Harness² method, and a Terminal-Bench audit that undercuts the benchmark's reliability.

Evidence

DeepSeek V4.1 Flash 1017. The AA Intelligence Index rose 6 points to 45, 1 point behind MiMo-V2.6-Pro. Other AA results:

• GDPval-AA Elo rose 146 points to 1746.
• CritPt went from 14% to 24%.
• The AA-Omniscience non-hallucination rate went from 4% to 33%.

On other leaderboards:

• Code Arena WebDev: #1 open and #12 overall at 1670, 52 points above the prior Flash.
• Agent Arena: #1 open and #11 overall, at a $0.20 median cost per task.
• Vals Index v2.2: #8 overall and the top open-weight model.
• Long-horizon browser use: 75/100, matching Opus 5.5, at $0.22 versus $2.35 per task.

The vendor cites Terminal-Bench 4.0 rising from 31.2 to 42.3.

MiniMax M3.1 Flash. AA scores it 42 at about $0.22 per task, level with GLM-5.3-Flash. HLE is 48% and AA-Omniscience accuracy is 56%. Its hallucination rate is 93%, because it attempts nearly every question. MiniMax's own claims are DeepSWE 74%, Automation-Bench 62%, and 1/3 the cost and 3× the speed of M3.

Harness² (Google Research, open source). An agent edits its own system prompt, skills, hooks and memory while the model weights stay frozen. The loop has three roles, and no grader is needed:

1. An editor agent proposes edits.
2. A composer diffs trajectories against a baseline.
3. A reflector stores lessons for later tasks.

With Gemini 3.5 Flash it adds 8.4 points on WorkBuddy-Bench and 7.5 on JobBench. About 50% of edits are skills and scripts.

vLLM v0.32.0. The release has 814 commits. New features:

• A /v1/systemone structured-decision endpoint.
• Low-precision KV cache across hardware, including NVFP4 and 4-bit KV on ROCm.
• A GPU n-gram speculator in Model Runner V2.
• Multi-API workers, reported at +43-85% req/s.

Separately, shipping precompiled Python bytecode in images cuts startup by 14-18% (about 18 s on H100).

How to use it

• DeepSeek V4.1 Flash 1017: available on DeepSeek's API, Fireworks, Baseten, DeepInfra, Cloudflare and Together.
• Local runs: Unsloth shipped llama.cpp support with dynamic GGUFs. The Engram/PLE tables can drop to Q6_K at roughly MXFP8 quality and can stream from disk.
• Community quants: 2-bit and 3-bit TrellisMX builds exist. The 3-bit build reports a mean KLD of 0.0664 nats.
• Throughput: a work-in-progress megakernel reaches 500+ tok/s on 4× RTX Pro 6000 Max-Q.
• MiniMax M3.1 Flash: costs $0.15/$0.60 per million tokens, with cache reads at $0.015. Weights arrive 23 October 2026. It is free for a limited time on OpenRouter, OpenCode, Cline and Ollama Cloud.

Limits

The DeepSeek gains are contested, and the evidence comes from several sources:

• AA's own caveat: MiMo-V2.6-Pro and GPT-6.1 Sol (medium) score higher at a lower cost per task.
• Zhihu tester: gains were larger in agentic execution than in reasoning. Exploration-heavy tasks used 30-40% more tokens with no better scores. The comparison also mixes in a DeepSeek Harness 0.2.1 change, not only the model update.
• @teortaxesTex: reports regressions on constraint adherence, browser use and research in one composite eval. The same commentator says it was the only model caught cheating on a bug-hunting eval.
• KernelBench: its #1 result came from storing the KV cache in FP8 against a BF16 spec.
• Deployment risk: the US Department of Homeland Security is ordering civilian agencies to remove DeepSeek and products built on it.

The Terminal-Bench audit (chem_safety) reviewed 65 tasks. It found issues in 60, cheating-passable tasks in 44, and grading on unstated requirements in 41. This weakens the vendor's Terminal-Bench 4.0 figures.

Most items are one-line tweet summaries with links. The page does not give papers, commands or full eval settings.