DailyWeights & Takes

22 Oct 2026 · the knobs come off

The last 24 hours, to 22 Oct 2026 10:15 ET

Edition 3 · 2 news · 2 model releases · 4 repos · 5 papers · 1 thing to try

At a glance

News 2

A short list today. OpenAI's CLI trended on HN, but it came out in May, so it isn't listed. Mistral Large 4 trended again too; its weights still aren't out.

  1. Anthropic

    Claude API version 2026-10-21, with new major SDK versions

    This is the first new anthropic-version since 2023-06-01, and the new SDK majors send it by default (Python 2.0.0, TypeScript 1.0.0, Go 2.0.0 and others). On the new version, a request that sends temperature, top_p or top_k fails with a 400, even at the default value. So does a document block in an assistant turn, or citations on user text. The SDKs also drop Text Completions, send keys as Authorization: Bearer, and move from MIT to Apache 2.0. The guide says a request that still needs a low temperature on an older model can stay on 2023-06-01. Claude Code 2.1.307 added /claude-api api-version to migrate your code for you.

    breaking changeplatforms release notesClaude Code 2.1.307

  2. Hacker News

    Omnara: an open-source, self-hostable alternative to Claude Managed Agents

    Omnara started as a phone remote for Claude Code. It now runs the agent loop on a server behind an API, with state in Postgres, so an agent survives restarts and costs nothing while idle. A small outbound daemon lets it use your laptop or other machines as tools, with no inbound ports. It works with any compatible model and self-hosts with Docker Compose; Managed Agents bills token rates plus $0.08 per active session-hour. Apache 2.0.

    94 pts27 commentsagents discussion

Model releases 2

  1. Recursive

    Flint: a 40B-total, 2.5B-active open-weight coding model

    This is the first model from Recursive, Richard Socher's lab. It's a mixture of experts built for coding agents, with tool use, reasoning and a 262K documented context. Recursive reports 54.9% on Terminal-Bench 2.1 and 74.2% on SWE-bench Verified, and more than 125 tokens/s on an M4 Pro MacBook Pro at short contexts. These are its own numbers; no independent leaderboard has run it yet. Recursive says its automated research system, the "Eureka Machine", chose the expert layout and the 4-bit output layer, and wrote a Metal kernel for M4/M5. Weights are Apache 2.0 (about 80 GB in BF16).

    422 likes on HFopen weightslocal model card

  2. Hugging Face blog

    IBM Granite Speech 5.0

    An open speech-recognition upgrade small enough to run yourself. The new 2B model reports 5.06% average English WER (second among open models on the Open ASR Leaderboard), plus translation and keyword lists of up to 2,000 terms. The 470M CTC model goes multilingual and is about 1.6× faster. The 4.1-2b-plus update improves speaker attribution and drops in without code changes. All Apache 2.0, with Transformers support.

    open weightsspeech

GitHub repos 4

All four broke out in the window: at least 50 stars a day and at least 3× their earlier rate. The repos created today were weak. Many were near-identical "genpark-*-skill" repos from two related accounts.

  1. GitHub

    lemomo-ai/lemo-opuscar

    59 film styles as a Claude Code plugin. The agent draws each frame in canvas and WebGL, with no video model, and adds music and narration. A film takes 30–60 minutes of agent time. The styles are tuned for Opus 5.5. MIT code.

    ≈ +1,040 ★ in windowwas ~171/daybreakout

  2. GitHub

    harry0703/MoneyPrinterTurbo

    This topic-to-short-video pipeline shipped v1.3.9 on 21 Oct. It now installs with pip as mpt, adds slow-zoom image motion, an optional local SDXL image server and more LLM providers.

    ≈ +893 ★ in windowwas ~140/daybreakout

  3. GitHub

    kuafuai/andee

    One Android APK that drives any app through the accessibility service, with no root and no per-app integrations. The agent loop runs on the phone, but the model is any OpenAI-compatible endpoint you point it at. The license adds limits on SaaS and embedding.

    +299 ★ since 22 Oct≈ 507/day · was ~62/daybreakout

  4. GitHub

    monid-ai/monid

    An "OpenRouter for agent tools": one key and three verbs (discover, inspect, run) over a catalog of paid tool APIs, such as search, enrichment and media generation. Discovery is free and calls are metered. The framework is MIT and can run locally with your own vendor keys.

    ≈ +376 ★ in windowwas ~96/daybreakout

Papers 5

Hugging Face daily papers featured on 22 Oct; submission dates shown.

  1. Hugging Face papers

    Cache Engineering for Agents

    The agent manages its own KV cache. It drops stale tool output and re-positions the context it keeps, without re-sending text or re-prefilling. A simple rolling drop of old tool responses scores 1.7 points higher and costs 27.7% less than summarize-when-full. The serving engine is open source.

    85 upvotesagents · inference

  2. Hugging Face papers

    Agent World Model in the Wild (WildAWM)

    Open models build synthetic desktops that mix websites, MCP services and local files, then check by execution that each task can be solved. Training Qwen3.6-35B-A3B only in those worlds adds 10.6 points on average across eight outside benchmarks, including +16 on OSWorld-Verified. Code, data and models are released.

    59 upvotescomputer use

  3. Hugging Face papers

    Harness Scaling for Generalizable Agentic RL

    The paper trains agents across 18 deliberately varied harnesses, with different prompts, tools, observations, context handling and workflows. Those agents do better in harnesses they never saw, including Claude Code, than agents trained in that black-box harness directly (41.2 vs 36.3 for Qwen3-30B-A3B). No code yet.

    12 upvotesagents · RL

  4. Hugging Face papers

    SpatialScout: can VLM agents explore and map 3D scenes?

    The agent has to explore a 3D scene and map every object, and frontier models still mostly can't. The best run, Opus 5.5, gets a 41.7% mean IoU and places only 3% of objects precisely, at about $29 a task. Models that triangulated more located objects better. Code.

    123 upvoteseval

  5. Hugging Face papers

    PrismaCoder: open multimodal code models

    These are 9B and 35B-A3B Qwen3.5 fine-tunes that turn charts, chemical diagrams and UI screenshots into code. The 35B version matches Claude Sonnet 4.6 on VisPlotBench and leads the open baselines on most suites. Models and data are on the project page.

    48 upvotesopen weights · code

Things to try 1

Flint in Ollama (Ollama 0.40.8, 21 Oct). A coding model that's open, fast and only 2.5B active parameters is the right size to test as a local backend for your agent. Run it on small, well-scoped tasks first, and compare its patches against your usual hosted model before trusting Recursive's benchmark numbers.

First step: update Ollama to 0.40.8 or later, then run ollama run flint. For now this works on macOS only; Ollama says support for other GPUs comes in a later release. On Apple Silicon, Recursive's own engine, rec, is the alternative (see its setup guide). On a single big NVIDIA GPU, use the SGLang kit from the model card.

The take

The knobs are coming off the panel. Settings you used to tune by hand are now chosen by the model, the provider or an automated researcher.

Anthropic's new API version rejects temperature, top_p and top_k outright; its newest models already accepted only the defaults. Recursive says its Eureka Machine chose Flint's expert layout and 4-bit output layer. Cache Engineering lets the agent decide which context to forget, and Harness Scaling trains models not to care which harness you wrap them in. When you can't tune something, you measure it, so your leverage moves to evals. Prediction: within a year, another major lab drops sampling parameters from its default API.