[AINews] not much happened today

What it claims

This Latent Space AINews issue covers 16 to 19 October 2026. Its lead item is DeepSeek V4.1 Flash 1017, which it calls the open-weight value leader. Artificial Analysis (AA) lists it as a 552B-total, 16B-active MoE with a 1M-token context window and an MIT licence. The issue also carries pricing, availability and eval disputes for several other releases, Google's Harness² method, vLLM v0.32.0 features and a Terminal-Bench audit.

Evidence

DeepSeek V4.1 Flash 1017

• AA Intelligence Index: 45, up 6 points. That ties GLM-5.3 for #2 open-weight, 1 point behind MiMo-V2.6-Pro, at $0.26 per task. AA notes that MiMo-V2.6-Pro and GPT-6.1 Sol (medium) both score higher at lower cost per task.
• GDPval-AA Elo: rose 146 to 1746.
• CritPt: rose from 14% to 24%.
• AA-Omniscience: the non-hallucination rate rose from 4% to 33%.
• Code Arena WebDev: #1 open and #12 overall with 1670 points.
• Agent Arena: #1 open, with a $0.20 median cost per task.
• Vals Index v2.2: #8 overall.
• Long-horizon browser use: 75/100, matching Opus 5.5, at $0.22 versus $2.35 per task.
• Terminal-Bench 4.0: the vendor figure rose from 31.2 to 42.3.

MiniMax M3.1 Flash

• AA scores it 42, at about $0.22 per task. 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%, BabyVision 91%, and 3× the speed of M3 at 1/3 the cost.
• Reddit commenters cite M3 as 428B total and about 23B active parameters. The issue does not state M3.1 Flash's size.

Mercury 3 (Inception) is a diffusion LLM. Inception claims it beats GPT-6 Luna and Haiku 5.5 on quality at 1,000+ tok/s. The issue gives no independent numbers for it.

Terminal-Bench audit. Of 65 tasks reviewed, 60 had issues, 44 could be passed by cheating and 41 graded on unstated requirements. This bears directly on the Terminal-Bench 4.0 figure DeepSeek cites.

How it works

Harness² (Google Research, Apache 2.0, repo google-research/harness2) lets an agent edit its own harness while the model weights stay frozen. The harness covers the system prompt, skills, hooks and memory. The loop has three roles:

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

No grader is needed. With Gemini 3.5 Flash it gains 8.4 points on WorkBuddy-Bench and 7.5 on JobBench. Skills and scripts make up about 50% of edits.

Muse Glimmer (Meta) is a 30B open-weight agentic model designed to run at 4-bit on one consumer GPU. Its recipe uses full-KL co-distillation from Muse Spark's in-training logits, then cross-entropy for the final 5% of training.

How to use it

DeepSeek V4.1 Flash 1017

• Pricing: $0.30/$1.20 per million input/output tokens at peak. Off-peak is half that. Cache hits cost $0.006 per million.
• Hosts: DeepSeek API, Fireworks, Baseten, DeepInfra, Cloudflare and Together.
• Local GGUFs: Unsloth's GGUFs (unsloth/DeepSeek-V4.1-Flash-1017-GGUF) need Unsloth's llama.cpp PR or Unsloth Desktop.
• Settings: recommended defaults are temperature=1.0, top_p=1.0, context=1M and max_tokens >= 256K.
• Chat template: none is shipped as Jinja, so users must use deepseek-recipe.
• Memory: Engram/PLE tables can drop to Q6_K at roughly MXFP8 quality and stream from disk.
• Reported runs: one user ran a 366GB ds4 build on a 128GB M5 Max at about 15 tok/s, with SSD fallback, and saw stable tool calling. A work-in-progress megakernel reaches 500+ tok/s on 4× RTX Pro 6000 Max-Q.

MiniMax M3.1 Flash

• Pricing: $0.15/$0.60 per million tokens, with cache reads at $0.015.
• Weights: due 23 October 2026.
• Free access: for a limited time on OpenRouter, OpenCode, Cline and Ollama Cloud.

vLLM v0.32.0 (814 commits) adds:

• a /v1/systemone structured-decision endpoint;
• NVFP4 and 4-bit KV cache on ROCm;
• a GPU n-gram speculator in Model Runner V2;
• multi-API workers, reported at +43 to 85% req/s.

Precompiled bytecode in images cuts startup by 14 to 18%.

Local fit: NVIDIA Nemotron 3.5 Bolt (74B total, 9B active, quantization-aware training to 2-bit) is the only recent large drop that fits in 24GB, at about 23.75GB.

Limits

• Harness confound. A Zhihu tester notes that DeepSeek's comparison also includes a change to DeepSeek Harness 0.2.1. Exploration-heavy tasks used 30 to 40% more tokens without better scores.
• Mixed composite results. @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 precision. Its #1 KernelBench result stored the KV cache in FP8 against a BF16 spec.
• US government use. DHS Order 2026-01 requires US civilian agencies to remove DeepSeek and products built on it. National-security systems are excluded, and agencies have 60 days to request exceptions. Commenters read it as targeting DeepSeek-operated hosting, not weights hosted elsewhere.
• Harness² caveats. Reddit commenters flag context-budget cost for local models and possible overfitting to benchmarks.
• Truncated source. Only 40,000 of 47,045 characters were available.