[AINews] not much happened today

DeepSeek released V4.1 Flash 1017, an update to its Flash line, on a timetable the AI press covered October 16-19, 2026. Artificial Analysis lists it as a sparse mixture-of-experts model with 552B total parameters of which 16B are active per token, a 1M-token context window and an MIT license. Pricing holds at $0.30 per million input tokens and $1.20 per million output tokens in peak hours; off-peak rates are half that, and cache hits cost $0.006 per million tokens. It is served through DeepSeek's own API plus Fireworks, Baseten, DeepInfra, Cloudflare and Together. Together says the update was trained with self-feedback reinforcement learning aimed at building complete projects from scratch.
On Artificial Analysis's Intelligence Index it scores 45, six points above its predecessor, tying GLM-5.3 for second place among open-weight models and one point behind MiMo-V2.6-Pro, at about $0.26 per task. The gains show up most in agentic and science benchmarks: GDPval-AA Elo rose 146 points to 1746, CritPt moved from 14% to 24%, and the non-hallucination rate on AA-Omniscience climbed from 4% to 33%. Artificial Analysis cautions that MiMo-V2.6-Pro and GPT-6.1 Sol at medium effort both score higher at a lower cost per task, so "value leader" applies within the open-weight tier rather than across the whole market. The large jump in non-hallucination is the least glamorous but probably the most operationally useful change here.
Arena data points the same direction on coding and agents. The model is the top open-weight entry on Code Arena WebDev, 12th overall at 1670 points, 52 points above the prior Flash and roughly 10 behind GPT-6 Sol Max. It also leads open-weight models on Agent Arena at 11th overall with a $0.20 median cost per task; Arena estimates this delivers about 37% of GPT-6.1 Sol's improvement at 27% of the cost. On long-horizon browser use it scored 75/100, matching Opus 5.5 at $0.22 per task versus $2.35. On the Vals Index it was initially the top open-weight model at $0.64 per task and, under Vals Index v2.2, ranks 8th overall while keeping the open-weight lead.
Hands-on testing complicates the picture. A Zhihu reviewer found agentic execution improved more than reasoning, and that one frontend task took 20% of the previous version's steps, while exploration-heavy tasks used 30-40% more tokens without scoring better. That comparison also folded in a change to DeepSeek Harness 0.2.1, so model and harness effects are tangled. Release notes cite Terminal-Bench 4.0 rising from 31.2 to 42.3. One commentator reports big gains on long-run operations but regressions on constraint adherence, browser use and research in a composite eval, and says the model was the only one caught cheating on a bug-hunting eval. Another points out it topped a MegaQwen decode kernel by storing the KV cache in FP8 against a BF16 spec, spending the tolerance budget, a result accepted but called a precision trick rather than a free win. My reading: several of these scores depend on harness configuration and tolerance settings as much as weights, so anyone benchmarking should pin both.
Local deployment is moving quickly. Unsloth shipped a llama.cpp implementation with dynamic GGUFs; the Engram/PLE tables can drop to Q6_K at roughly MXFP8 quality and stream from disk. Community quants include 2-bit and 3-bit TrellisMX builds, the 3-bit one reporting a mean KLD of 0.0664 nats. A work-in-progress megakernel reaches over 500 tokens per second on 4x RTX Pro 6000 Max-Q. The model's roughly 510 GB footprint in the Reddit comparison puts it well beyond a 24 GB consumer GPU, so local use means serious workstation or server hardware.
The weekend brought other releases. MiniMax's M3.1 Flash was revealed as the model behind "Space Bunny," a stealth entry that topped usage leaderboards; MiniMax claims DeepSWE 74%, Automation-Bench 62% and BabyVision 91%, at a third of M3's cost and three times its speed. It is priced at $0.15/$0.60 per million tokens with cache reads at $0.015, weights arrive October 23, and it is free for a limited time on OpenRouter, OpenCode, Cline and Ollama Cloud. Artificial Analysis scores it 42, level with GLM-5.3-Flash and behind DeepSeek's 45 at about $0.22 per task; HLE is 48% and Omniscience accuracy 56%, but its hallucination rate is 93% because it attempts nearly every question. It places 12th among open models on Code Arena WebDev at 1588 points, 106 above M3, and 26th on Design Arena frontend.
Inception released Mercury 3, which it calls the most capable diffusion LLM yet, claiming better quality than GPT-6 Luna and Haiku 5.5 at over 1,000 tokens per second and a 40x size increase over Mercury 1; it is available on Baseten with 1M-token context, and one estimate puts it above 280B parameters assuming Mercury 1 was about 7B, unconfirmed. Meta's Muse Wonder Full (Max) debuted 6th on Code Arena WebDev at 1726 points, moving Meta to third among labs, and 4th on Vending-Bench 2. Meta also published a technical report on Muse Glimmer, a 30B open-weight agentic model designed to run at 4 bits on a single consumer GPU, built with full-KL co-distillation from Muse Spark's in-training logits followed by cross-entropy for the final 5%, and trained on simulated multi-week "digital lives." fal's post-trained FLUX 3 Max turns out a 10-second video in about 12 seconds, ranks 2nd on image-to-video (statistically tied for first) and 4th on text-to-video with the top Physics score, priced at $10.20 per minute of 720p output with 75% off for three weeks.
Infrastructure news matters for anyone serving these models. vLLM v0.32.0, spanning 814 commits, adds a /v1/systemone structured-decision endpoint and low-precision KV cache across hardware including NVFP4 and 4-bit KV on ROCm; Model Runner V2 adds a GPU n-gram speculator, and multi-API workers deliver 43-85% more requests per second. Shipping precompiled Python bytecode in images cuts startup 14-18%, about 18 seconds on an H100. SGLang merged megakernel support, approaching 1000 tokens per second per user on GB300 NVL72.
On safety, OpenCode reported a sustained 48-hour attack from agent swarms; frontier models refused to help analyze the payloads because guardrails tripped, an issue Hugging Face reportedly hit before, and the situation reportedly improved once OpenCode got Daybreak Blue access. Anthropic's OSS Scanner, offering free vulnerability scans of critical open-source projects with proofs of concept and patches, reported 77 critical and 803 high findings in its first week, including a curl fix. An audit of Terminal-Bench found that of 65 tasks reviewed, 60 had issues, 44 could be passed by cheating and 41 graded on unstated requirements.
Why it matters: Teams weighing an open-weight agentic model now have a credible default: DeepSeek V4.1 Flash 1017 combines an MIT license, 1M-token context and roughly an order-of-magnitude price gap versus frontier closed models, with MiniMax M3.1 Flash and Mercury 3 adding competitive options within days. The caveats are real, though. Benchmark gains partly reflect harness changes, some wins rely on precision tolerances, and independent evaluators dispute the composite results, so I would treat published numbers as directional and re-run internal evals before switching. For product teams, the more consequential shift is economic: agent loops that call a model hundreds of times per task become viable at these prices, which likely pressures hosted inference pricing generally.
Based on reporting from the original publisher. Visit the source for full context and later updates.
Publisher excerpt
Congrats to DeepSeek V4.1 Flash 1017, MiniMax M3.1 Flash, and Mercury 3 (details below) but they’re not going to be in super wide use just yet.