AI Daily · Oct 22, 2026: Recursive releases Flint, a 40B open-weight coding model, free on OpenRouter
Recursive open-sourced Flint, a 40B MoE coding model with ~2.5B active parameters and 262K context, free on OpenRouter and running 125+ tok/s on a MacBook Pro; MiniMax M3.1 Flash also launched, free on AI/ML API.
Pick a date
-
Recursive releases Flint, a 40B open-weight coding model, on OpenRouter
Flint is Recursive's first model, an open-weight 40B MoE with ~2.5B active parameters, 262K context, reasoning and tool calling, built by its Eureka Machine system. It reports 54.9% on Terminal-Bench 2.1, 89.7% on SWE-bench Pro v2 and 74.2% on SWE-bench Verified, and is free to use on OpenRouter.
-
MiniMax launches M3.1 Flash model with launch partners
MiniMax released M3.1 Flash, described as faster and lighter than the earlier Space Bunny version that accumulated over 200T tokens, with partners onboard from day one.
-
Malicious LiteLLM 1.107.0 artifacts published to PyPI and Docker Hub
LiteLLM reported that malicious 1.107.0 artifacts were published to PyPI, Docker Hub and GHCR around 4:40 a.m. PDT and detected at 6:32 a.m. PDT. Enterprise deployments pulling only from its private registry are unaffected.
-
Claude Security adds continuous code re-scanning in public beta
Claude Security now runs a deep first scan after connecting repositories, then re-scans as code changes and marks findings fixed or reopens them. It is in public beta on Claude Enterprise, spans GitHub/GitLab repos, and eligible orgs get $25K monthly usage credits for two months through Dec 31, 2026.
-
Perplexity introduces JOLT post-training method matching GRPO with 6.5x fewer tokens
Perplexity's JOLT lets one model act as both teacher and student, with the teacher also trained on outcome rewards and penalized when its predictions diverge from the student's. On MATH it matched GRPO accuracy using 6.5x fewer training tokens, and JOLT+ beat GRPO on all five benchmarks across math, coding, tool use and terminal tasks, including a 9.9-point gain on AppWorld.
-
Anthropic opens research preview of Claude Duet proactive agent app
Claude Duet is a new app where Claude proactively surfaces items needing attention and completes tasks while the user is away, checking before acting. It is an early research preview available via waitlist to Pro and Max subscribers in the US, and a new classifier blocks sharing of sensitive data such as passwords.
-
Carbon-X 14B open DNA model released, trained on 300B tokens
Carbon-X is a 14B model mid-trained on 300B tokens of interleaved DNA and text, then SFT'd; it beats all open language models tested on DNA benchmarks and comes within a point of GPT-6.1 Sol at over 1,000x lower inference cost.
-
SGLang-TorchTPU plugin brings top open models to Google TPU
SGLang and Google's TPU team released SGLang-TorchTPU, a TPU plugin supporting Muse Spark Open, GLM-5.3 Flash, Kimi K3, Ling 3.1 Flash and MiniMax H3, and it is already used for production inference at Meta.
-
GMI Cloud offers 40% off DeepSeek V4.1 Flash for 10 days
GMI Cloud is discounting DeepSeek V4.1 Flash 1017 by 40% for 10 days with unlimited usage through its API.
-
Codex usage on Pro plan no longer counts toward weekly quota
OpenAI's Codex is fully included in the Pro plan and does not draw on the weekly usage allowance.