AI News - 2026-10-20

DeepSeek opens its model-serving software

• Continued: DeepSeek publishes the software that runs its model — day 3 of coverage (base specs in yesterday’s digest).
  • The Apache-2.0 code supports DeepSeek-V4.1-Flash, constrained JSON output and overlapping CPU scheduling with device work. It includes offline execution examples and performance-test scripts.
  • Supported hardware is NVIDIA SM100 GPUs or Huawei Ascend 950 processors. Prerequisites include Linux, GCC 13 or newer and Python 3.12, plus CUDA 13 or CANN 9.2 respectively.
  • Published tests use four SM100 GPUs for prefill, the model’s first pass over a prompt, before it writes anything. Answer generation uses 64 GPUs in one tightly connected group.
  • A complete service still needs infrastructure for the KV cache, a model’s stored working memory for a prompt. Request routing and separation of prompt processing from answer generation also remain integration work.
  • Reproducible sampling requires a fixed per-sequence seed, batch-invariant mode and disabled variable-length drafting. Batch-invariant mode is NVIDIA-only.
  • Caveat: Building a hosted service still requires additional infrastructure, and the supported hardware excludes ordinary consumer cards. (HN 328 points · 110 comments)

Agent frameworks & tooling

• Cloudflare’s Clairic pulls typed values from documents and images — Its hosted interface offers a cheaper component to test for extraction steps that currently use a general chatbot.

  • The preview API charges $0.10 per million input tokens, with no output-token charge. Requests accept text and up to four images.
  • Text fields return source positions and confidence scores. Numeric estimates return a median and an 80% interval, giving downstream code information for routing uncertain cases to review.
  • Cloudflare’s damaged-product example took 19.3 seconds and cost $6,032 per million claims with Clairic, Clef and GLM-5.3. Using GLM-5.3 for every step took 61.7 seconds and cost $11,740.
  • That comparison covers a small vendor demonstration. The model struggles with longer arithmetic chains, and its uncertainty ranges can become too narrow on unfamiliar data. Downloadable weights remain planned.
  • Caveat: Valid types and confidence scores do not guarantee correct values, so financial calculations and consequential decisions still need independent checks. (Techmeme · lobste.rs)

• Combining two instruction files can make a model ignore one paper notes → — SkillClash supplies a practical test for overlapping skills and a simple repair worth trying before adding more instructions.

  • An agent is a model running a multi-step task with tools. This study places both skill files directly in context, removing file selection as a cause of failure.
  • Across 16 constructed skill pairs and seven models from five families, rule compliance fell from 93–99% with one skill to 61–62% with two.
  • A merged file written by Claude Opus 5.5 restored 96–99% compliance with achievable rules across all seven models. Its correct worked example supplied much of the gain.
  • Adding an example that follows both skills to each file reduced complete takeover to 1–15% across six tested models. The code and recorded outputs are available.
  • Caveat: The study measures rule compliance on constructed text tasks, with real-repository work left for separate evaluation. (arXiv 2610.25064)

• Scale and Reflection explain how they repaired coding-agent tests paper notes → — Their October 16 paper shows why trustworthy comparisons need clean instructions, restricted network access and fresh-container grading.

  • A benchmark is a fixed set of tasks for comparing models. Engineers repaired or retired 480 of the original 731 tasks, leaving 642, with a 51-task Hard subset.
  • Twenty-three engineers spent 404 hours on review. A second engineer checked each revised instruction by solving it without seeing the reference solution.
  • The protocol permits only the model endpoint during agent work, then applies the submitted patch to a clean container. This prevents unrelated sandbox changes from earning credit.
  • The dataset dates V2 availability to September 22. Reported scores reach 99.4% on the full set and range from 52.9% to 98.0% on Hard.
  • Caveat: V2 is already near saturation for the strongest tested systems, so a high score offers limited evidence about harder or unfamiliar engineering work. (arXiv 2610.24206)

Models & research

• MiniMax releases M3.1 Flash at lower hosted prices — It provides another inexpensive coding and computer-use model to compare against your current provider.

  • Pricing is $0.15 per million input tokens, $0.60 per million output tokens and $0.015 per million reused input tokens.
  • MiniMax reports 73.7% on DeepSWE v1.1 coding tasks, 61.7 points on OSWorld V2 computer tasks and 29.6 points on Agents’ Last Exam professional tasks.
  • The company reports median output speed above 170 tokens per second, with some requests exceeding 200. Its claimed 70% task-cost reduction compares against its predecessor.
  • Access is available through the MiniMax API, M Plan and MiniMax Code. Model weights are promised for later release.
  • Caveat: These are MiniMax’s own measurements, and current access sends your requests to a hosted service. (HN 472 points · 200 comments)

• Inception releases Mercury 3 with a million-token context — Its text-generation method refines multiple tokens together, making it worth testing when model waiting time slows a multi-step workflow.

  • The release announcement reports 1,068 tokens per second on NVIDIA GPUs. Its claimed four- to eightfold speed advantage depends on the compared models and test setup.
  • The API documentation supports a one-million-token context and up to 131,072 output tokens, with tool requests and schema-constrained answers.
  • Launch prices are $0.10 per million input tokens, $0.01 for reused input and $0.35 for output. These are 50% discounts from $0.20, $0.02 and $0.70 respectively.
  • It uses the OpenAI-compatible chat-completions interface. The launch discount has no stated end date in the collected reporting.
  • Caveat: Fast token generation alone does not establish faster successful tasks, and the launch prices are promotional. (Techmeme)

• Automatic prompt tuning helps these local models with math paper notes → — The two-model study gives DSPy users a concrete reason to keep untouched test questions separate from optimization feedback.

  • GEPA revises prompts using feedback from earlier runs. On 250 math questions per condition, exact-answer accuracy rose from 82.8% to 90.4% for Qwen3.8-27B and from 85.2% to 90.8% for DeepSeek V4 Flash.
  • On 500 HotpotQA questions per condition, neither optimizer improved the baseline. Qwen’s SkillOpt development score rose from 62.0% to 77.6%, while its separate test F1 fell from 82.15% to 78.93%. F1 measures overlap with expected answer words.
  • Qwen’s math responses averaged 99.35 seconds with GEPA versus 143.34 seconds without it. Total tokens per successful response rose from 1,382 to 2,253, about 63%, because the prompt grew.
  • Target models ran locally, but Claude Opus 4.8 supplied hosted optimization feedback. Public supporting files include prompts and results, while the authors’ benchmark implementation is omitted.
  • Caveat: Each model’s conditions ran sequentially, so observed speed changes remain sensitive to serving behavior as well as prompt quality. (arXiv 2610.26020)

Industry

• OpenAI releases tests for coding in less-common languages — MultiLangBench gives teams a repeatable check before trusting model-generated changes in older systems.
  • The Apache-2.0 repository asks models to implement SQLite 3.40.1. It converts about 24,000 of SQLite’s 51,000 tests and recommends a 272-test subset for ordinary comparisons.
  • Models receive documentation but no internet access. Existing SQLite copies must be removed, and grading checks the resulting database with real SQLite.
  • OpenAI reports average test pass rates across 20 languages of 87.4% for GPT-6 Astra, 40.3% for GPT-6 Luna and 26.8% for GPT-5.6 Sol.
  • A separate COBOL migration test requires exact observable outputs. GPT-6.1 Sol scores 86.6% at low reasoning effort versus 5.4% for GPT-5.6 Sol, at 39% lower API cost; higher effort reaches 97.9%.
  • Caveat: Passing these controlled tests is different from preserving every business rule and dependency in a production migration. (Techmeme)

Policy & provenance

• WhatsApp hid instructions discouraging chatbot users from contacting regulators — The finding shows why retrieved help text must remain evidence rather than instructions controlling an assistant’s advice.

  • Redact’s investigation saved 545 help pages and found 102 hidden blocks on 80 pages. Most blocks contained ordinary housekeeping guidance.
  • Two blocks told the chatbot to avoid recommending a regulator after account-deletion trouble, or citing privacy laws once lawyers or litigation were mentioned.
  • The notes were intended for WhatsApp’s support chatbot. Servers also sent them to Google and Bing crawlers; the investigators’ requests under several AI-crawler names returned neither article nor notes.
  • The notes have been removed from live pages. The report links October 9 and October 18 archived copies for comparing the deletion-page wording.
  • Caveat: The researchers inspected instructions but did not test whether WhatsApp’s chatbot or outside assistants actually followed them. (HN 466 points · 137 comments)

• Anthropic’s draft would replace some state AI oversight rules — The proposal matters because it would shift specified reporting and audit duties from states to a federal framework.

  • The linked 37-page discussion draft is marked September 29. It would preempt state disclosure, evaluation and incident-reporting duties for covered risks, while preserving specified powers including child-safety rules and common-law claims.
  • Commerce would need a court order to restrict a model. The court would decide the required questions without deferring to Commerce’s determinations.
  • The linked copy requires evaluator registration and a rating system within 180 days. That differs from Transformer’s description of evaluators simply self-certifying without standards.
  • The article reports that Anthropic’s federal-affairs head promised revisions to imprecise preemption language. The linked copy’s operative preemption list also omits its separately defined model-suspension category.
  • Caveat: This is an unsettled discussion draft, so it changes no law today and its exact version matters. (Techmeme)

All gathered items - what was cut and why (6)

• Google’s Harness² - STALE: the paper dates to October 9, and repository history shows its release commit on October 8. A new social post supplies no new checked result. (X @_philschmid)
• Maxwell’s program acceleration - LOW_UTILITY: the site has concrete demonstrations, but access requires approval and the sweeping correctness claims lack an inspected public implementation. (HN · Maxwell)
• PLACE’s memory controls - LOW_UTILITY: relevant method, but its primary test uses one constructed 302-turn history and trusted event/query meanings. Released skill tests take priority today. (arXiv)
• Lobste.rs’ minimal-agent study - STALE: its sole submission is September 30, with no checked new experiment or revision. (lobste.rs · arXiv)
• DeepSeek’s reported federal-agency ban - UNVERIFIABLE: direct extraction was blocked, leaving the order’s exact scope and exceptions unchecked. (Reuters)
• Websites restricting assistant access - DEDUP: the restrictions were already covered; today’s collected post adds no verified change. (Bluesky · The Verge)