MiMo v2.6's Live Training Feed, Grok 4.7's Price War, CI Bottleneck

Good morning. It's Tuesday, September 22, and we're covering a rare open-weights release with a live training feed, a pricing move aimed at the most expensive model in the market, and a CI rework every agent-heavy team will recognize.

Two frontier-class model announcements landed within hours of each other yesterday, but the story with longer legs is how Xiaomi shipped its: the MiMo v2.6 thread sits at 912 points with 406 comments, the live post-training dashboard hovered near the top of HN all day, and the Grok 4.7 thread pulled 488 comments of its own. Here is the signal, sourced.

In today's brief:

• Xiaomi released MiMo v2.6 with MIT-licensed weights: a 1.02T-parameter sparse MoE (42B active) that trained post-RL in front of a live public dashboard, complete with real-time loss, benchmark scores per checkpoint, and running cost estimates
• SpaceXAI shipped Grok 4.7 at $2/$6 per million tokens - a fifth of Fable 5.1's output price - with a 46.3% CursorBench 4.0 score and a faster variant at double the output speed
• Linear published why AI coding made CI the bottleneck at their scale - fourfold test growth, 87,000 runner-minutes a month saved - and how they reworked it
• Cloudflare declared Python Workers generally available with native bindings, FastAPI/Django/Flask connectors, and databases over socket syscalls in Wasm

THE BIG ONE

MiMo v2.6: A 1T-Parameter Open Model That Trained Its RL Run in Public

Xiaomi's MiMo team released MiMo v2.6 (HN thread, 912 points, 406 comments), and the numbers alone would have carried the day: the flagship MiMo-V2.6-Pro-RL is a sparse MoE with 1.02T total parameters, 42B activated, a 1M-token context window, native text, image, video, and audio input, and an MIT license that clears every commercial-use question at once. On DeepSWE v1.1 it scores 71.9 against 74.0 for Claude Opus 5, 73.0 for GPT-5.6 Sol, and 70.0 for Fable 5 - the top open-weights score on that benchmark so far, with a smaller Flash sibling at 67.9 and day-zero GGUF and MLX quantizations appearing within hours.

The architecture is a study in how far the open field has moved since last year: "You Only RL Once" runs one mixed RL pass across coding, general agents, visual work, and cybersecurity instead of per-domain runs; a groupwise agentic grader (GRS + GAR) ranks passing rollouts against each other rather than against a binary pass/fail label; and MOPD2 handles multi-prefix distillation into hard-to-verify tasks. The technical report is unusually candid about the machinery, including anti-reward-hacking verifier cross-checks. Artificial Analysis measured the Pro at ~125 output tokens per second on Xiaomi's own API at $0.43/$0.87 per million tokens - about half the open-weights median on output pricing.

But the detail the HN thread kept circling back to is not the benchmark table - it is that Xiaomi ran the entire RL post-training phase behind a live public dashboard, the same one discussed on HN a few days earlier (item 49732270, 155 comments). Commenters logged what it showed: real-time training loss, benchmark scores after every checkpoint, running compute-cost estimates, and each restart logged with a reason - including one GPU OOM triggered by expert load imbalance mid-run, and a decision to drop a cyber dataset after it degraded coding benchmarks. One commenter described it as "an incredible learning and teaching tool," and several noted the on-policy RL telemetry - rollout-model staleness and a KL-based train-infer divergence - is the kind of thing labs normally keep private. The transparency is the point, and Anthropic's anti-open-weights arguments look thinner when a competitor publishes its RL loss curves.

Why it matters: MIT-licensed weights within spitting distance of Opus 5 on agentic coding reset the local and self-hosted bar - our local coding LLM roundup and open-weights economics analysis are the frames for deciding where it fits.

PLATFORMS

Grok 4.7 Prices the Frontier at $2/$6 and Keeps the Speed

SpaceXAI announced Grok 4.7 (HN thread, 576 points, 488 comments), its most coding-capable model yet, with a pricing move that does the talking: $2 per million input tokens and $6 per million output, versus $10/$50 for Fable 5.1 and $4/$20 for GPT-5.6 Sol - a fifth of the input price and roughly an eighth of the output price. The headline overhead is "twice as fast at half the price of comparable models" - a fast variant serves double the output speed at double the price - and the announcement chart puts Grok 4.7 at the price-performance frontier on CursorBench 4.0, where it scores 46.3% against Fable 5.1's 51.8% and GPT-5.6 Sol's 41.7%.

Under the hood it is a larger base model than Grok 4.6, trained with a longer RL run weighted toward tasks that take hours rather than minutes, with what the team calls its best-calibrated safeguards: 3.3% of risky dual-use prompts pass on the internal HackerBench v0.3, and it tops LatchBio's biosafety benchmark. The benchmarks are a mixed picture - Terminal-Bench 4.0 trails Fable 5.1 badly (38.0% vs 57.9%), while Harvey's legal agent benchmark leads the field at 19.6% and DeepSWE lands at 71.0 in high-effort mode. For teams already running routers the interesting part: a frontier-class model at one-eighth the output price of the field's most expensive flagship shifts routing economics noticeably, and our model routing breakdown and pricing comparison are updated frames for the trade.

Why it matters: the second consecutive cycle where the most expensive flagship had to price-match a value play changes the cost-per-task math for every coding agent deployment - the pattern we first tracked in the Fable 5 cost-per-task analysis.

PLATFORMS

Linear: Agents Made CI the Bottleneck, So We Reworked It

Linear's CEO Mufeez Amjad published the engineering post of the day, "AI coding has made CI a bottleneck" (HN thread, 246 points, 271 comments), and the numbers describe the pressure every agent-heavy team is starting to feel: test suites have nearly quadrupled since the start of the year, the codebase gains roughly 2,000 tests a week, and agents ship code faster than validation can accept it. The rework brought PR wait from over 6 minutes to just over 5, and cut runner time per test roughly in half, against a suite that kept growing.

The post is a catalog of compounding micro-fixes, most of them transferable: moving off GitHub-hosted runners to faster third-party machines (34% faster jobs in an A/B), switching to the native tsgo compiler (weekly median tsc check down 73%), and rewriting type-aware lint rules as pure AST checks so ESLint drops TypeScript entirely (API lint down 68%). Gating jobs got the biggest relative wins: capped-fetch checkouts took the slowest change-detection gate from 94 to 20 seconds, a custom checkout action with backoff and GIT_HTTP_LOW_SPEED_LIMIT killed the hangs, and moving cache-marker writes off the merge critical path saved 42 seconds per API PR. Database setup went from replaying migrations to loading a schema snapshot, cutting 12 seconds to 1-2 per container.

The largest single win is the most counterintuitive: Vitest's default isolate: true was rebuilding the entity, GraphQL, and decorator graph in every test shard, so Linear added an opt-in project with isolate: false that shares the module registry inside each worker - worth roughly 17% of monthly savings, with the slowest shard dropping from ~300-380 seconds to ~195. Files opt in with a comment, teardown was handled, and agent skills were updated so generated tests follow the same constraints.

Why it matters: "agents make review and CI the bottleneck" is becoming the defining infrastructure problem of agent-heavy engineering - this is the most concrete published playbook we have seen, and it lines up with the review-queue pressure in AI coding agents review queues.

PLATFORMS

Python Workers Go GA With Frameworks, Databases, and Agents

Cloudflare declared Python Workers generally available (HN thread, 229 points), two years after the beta launched - the language is first-class on the platform rather than a Pyodide guest. The headline additions: native bindings to D1, R2, Queues, Durable Objects, Hyperdrive, and Workers AI with no JS glue code - the pyodide.ffi.to_js dance is gone - plus workers.asgi and workers.wsgi connectors that run FastAPI, Django, or Flask apps through the Workers runtime as if uvicorn or gunicorn were handling them.

The two engineering details that make the platform credible for real workloads both solve Wasm's hard problems. TCP sockets were implemented at the syscall level by routing standard socket calls through the Workers connect API, which makes Postgres and MySQL drivers like aiomysql work at all - impossible before - and that is the same mechanism behind Hyperdrive; the socket bridge also unblocks HTTP clients, so openai, langchain, and the official Python mcp package run natively. On the ecosystem side, PEP 783 - which standardizes the PyEmscripten platform for Python-in-browser runtimes - was accepted after over a year of work, and cibuildwheel gained PyEmscripten support, so package maintainers can publish WebAssembly wheels once and have them work across every PyEmscripten environment, Cloudflare included.

Why it matters: Python Workers closing the socket and packaging gaps makes edge Python real for database-backed APIs and agent workloads, and it is the clearest signal yet that Cloudflare is building toward a full agent platform - the ADLC direction we covered.

RESEARCH

Decision Models: TypeSafe's Jev Returns Floats Instead of Tokens

Simon Willison's writeup of Jev, TypeSafe AI's new "System One" model category, is the most thought-provoking model story of the week. Jev accepts text and structured state but returns floating-point numbers rather than text: yes/no confidence (Noul questions, short for Bernoulli), choice questions with a probability distribution across the options, and score questions over a described range. It is deliberately cheap - $0.042 per million input tokens, output free, cheaper than OpenAI's GPT-5 Nano - and questions within one prompt evaluate in parallel, so scoring a whole candidate set costs milliseconds and a few cents.

Willison positions it as a decision model rather than a language model - spam flags, labels, prioritization, and his own experiment of BM25-fetching 100 matches and having Jev rank them against a query. He is uneasy about it in the right places: you get a float back and no justification, which makes bias auditing harder, and he flags the risk of ranking people. The open-weight community is already circling the shape - Kev rebuilds the format on Qwen 3.5 at 0.8B/4B/9B, a JevBench benchmark has appeared, and a 2048-playing gist is circulating.

Why it matters: if the float-output model shape holds for classification and reranking, a meaningful slice of LLM budget - the "is this spam/is this relevant/rank these 100" workload - stops needing token-by-token generation entirely.

SECURITY

Spymarks, Not Watermarks: The Tracking Payload Hidden in Provenance

Brandon Thomas's Spymarks, not Watermarks essay (HN thread, 452 points, 113 comments) sharpens a distinction that matters for anyone shipping or consuming provenance-marked AI content: the difference between a visible ownership watermark and a hidden signal that makes your work traceable without your knowledge. The technical anchor is Google's own SynthID-Image paper (arXiv 2510.09263), which reports the SynthID-O variant can encode a 136-bit payload in a 512x512 image - enough for a 64-bit database identifier with 72 bits of error correction left over, and those identifiers can map to user records: names, IPs, dates of birth, addresses.

That framing reframes the provenance conversation. Text spymarks steer word choices into a detectable statistical pattern; audio spymarks modify waveforms or frequency features and survive compression; the open-source audiowmark tool has done 128-bit AES-protected payloads since 2018. The point is not that every implementation is malicious, but that "watermark" now covers both a visible Getty-style overlay and an invisible identity-bearer that survives editing and metadata stripping. The useful test it proposes: EXIF and ID3 tags are standardized, inspectable, and editable; spymark payloads are none of those things. For labs and the tools that consume their output, the open question is whether provenance infrastructure means verification (whose model made this) or surveillance (which individual made this), and the default answer is increasingly the latter.

Why it matters: if provenance signals carry account-linked identifiers, "audit where your model output and content pipeline emit markers" becomes a privacy decision, not a compliance checkbox - an extension of the agent identity and security layer concerns we have tracked all year.

TOOLS WORTH A LOOK

1. Transformer Explainer (free, open) - Georgia Tech's Polo Club shipped a visual, click-through explanation of the transformer architecture from attention to layer norms (420 points on HN).
2. dlab Open Source Week (OSS, free) - Tim Dettmers' lab releases two projects and four papers as one ecosystem: quantized local inference that runs a 550B DeepSeek V4.1 on a 128GB MacBook, a CliffCompaction auto-compaction technique he claims beats Claude Code and Codex, and a fully local autonomous research system (161 points on HN).
3. Can gzip be a language model? (free) - Nathan's experiments pushing compression-based inference where it has no right to work; the reward is a better intuition for what language models actually exploit (151 points on HN).
4. MiMo v2.6 Flash GGUF (free, open) - day-zero llama.cpp quantizations of the Flash variant for anyone who wants to try a 67.9-point DeepSWE model locally before committing to the full 1T flagship.

WHAT ELSE IS HAPPENING

• "I don't want to read what you didn't write" (697 points): Colin Breck on the revolt against AI-written prose, citing a survey where 78% of developers stop reading posts they suspect are AI-written (7 min read).
• What Sun got wrong (594 points): Bryan Cantrill on Sun Microsystems and the engineering-company mistakes that still echo (essay).
• Attention is all you have (818 points): Alice Girard Guittard on attention hijacking and rebuilding an intentional internet of bookmarks and RSS (8 min read).
• Amazon blocks Meta's Muse agent from shopping (147 points): the retail side of agent egress policy - Amazon is refusing Meta's shopping agent at the door; our Meta Muse coverage has the model context.
• Git 2.56 and 3.0 ahead (131 points): LWN's preview covers the planned bump to Git 3.0 and what gets deprecated on the way (2 min read).
• Claude status: elevated errors (115 points): multiple Anthropic models served elevated error rates Monday evening; agent fleets should check retry and backoff behavior against the incident window.
• Fable 5 median thinking declined in August (391 points): a widely shared thread reports Fable 5's median thinking time fell through August - readers are split on whether it signals efficiency or evasive truncation; our Fable 5 usage-limits breakdown is the reference frame.

FROM THE SITE

What We Published Since the Last Brief

Put Your Coding Agent in Discord: A Team Ask-Bot Built in an Hour - the applied playbook for turning a Discord channel into a coding-agent front door, plus its three companions: agent behind a webhook, agent on a cron job, and Slack Code vs Claude Code vs Cursor vs OpenCode for when agents belong in chat at all.

Agent Retrieval Bench Finds The Files Before The Fix - why the new benchmark separates context acquisition from final answer, and what it says about agent memory and retrieval budgets.

---

Every link above goes to a primary source or our sourced coverage. Tomorrow's brief lands when the news does - subscribe to get it by email.