Three launches in four days, all priced like a rounding error, all with benchmark tables, and a front-page thread full of people asking who checks them. Plus: a 209M-parameter forecaster that runs on a laptop CPU, a diffusion LLM that claims 1,000+ tokens a second, and a Google Research harness that rewrites itself between tries.

Today: Space Bunny unmasked, Mistral’s time-series model, Mercury 3, recursive harnessing, a site that audits the labs’ own scores, and the Meta engineer who said what everyone was thinking.

Every lab says it is first and cheapest. The interesting question this week is who checks.

🐰 Space Bunny gets a name

For a while, people using OpenCode Go got a free model called "Space Bunny." Nobody was told what it was. This morning MiniMax said it was an early version of M3.1 Flash and that it had processed more than 200 trillion tokens while anonymous. On Hacker News, a commenter called DCKing was unimpressed: it was served free for a while, they wrote, and "failed to impress anyone as far as I can tell." Another, mawadev, asked the question under the whole thread: is anyone independently verifying any of this?

MiniMax’s own numbers: 73.7% on DeepSWE v1.1, 61.7 on OSWorld V2, 82% on CyberGym, 200+ output tokens a second (median 170+), and a flat $0.15 per million input tokens, $0.60 output, $0.015 cached input. They say that is 70% cheaper than the previous model for agentic tasks. All of that is the vendor’s table, run by the vendor. The post says the model is on the MiniMax API, M Plan and MiniMax Code, and that weights "will be released later," with no date.

MiniMax announcement  ·  HN discussion

📈 A forecaster that fits on your CPU

A commenter on the Timestral thread, logancbrown, put the real use case plainly: you could train an XGBoost model on one series with holiday covariates, but what if you have 10,000 different series to forecast? Ten thousand separate models each have little history. One generic model has to cope with everything. That gap is where a pre-trained forecaster is supposed to help. Another commenter, davedx, said they were migrating their macro team to it. Others, like beernet, called the whole idea marketing and noted Mistral’s post has an imaginary customer, not a real one.

Mistral says its largest model, 209M parameters, ranks first on three public forecasting benchmarks, GIFT-Eval, TIME and fev-bench, among pre-trained zero-shot models, from a single checkpoint with no ensembling. The family has four sizes. It handles multivariate series and both past and known-future covariates (production schedules, weather forecasts). Mistral’s CPU numbers: on 8 threads, forecasting one day ahead from three weeks of 15-minute history takes 10 ms with the 5M model and 140 ms with the 209M one. Those are Mistral’s measurements. The mistralai/timestral repo, created October 15, had 582 stars and an Apache-2.0 license on the repository this morning. We read the post and the repo page; we ran nothing.

Mistral announcement  ·  The repository  ·  HN discussion

⚡ A language model that drafts in parallel

Inception’s co-founder and chief scientist, Volodymyr, showed up in the HN thread himself as volodia: Mercury 3 is "our most capable production diffusion LLM." Simon Willison, in the same thread, wondered why Google did "effectively nothing" with last year’s Gemini Diffusion demo, then drew pelicans with it through OpenRouter and the Inception API and called them "a little quirky."

A diffusion LLM refines a whole draft in parallel instead of writing one token at a time. Inception says Mercury 3 outperforms GPT-6 Luna and Claude Haiku 5.5 on a range of benchmarks at matched reasoning effort, runs 1,000+ tokens a second (1,068 on NVIDIA Blackwell, in their words) which is 4 to 8 times faster, and holds 1M tokens of context, up from 260K. Price: $0.20 per million input and $0.70 per million output, with 50% off at launch. All speed and quality claims are Inception’s own. Simon’s pelicans are the only independent output in the thread, and they are an SVG toy, not an eval.

Inception announcement  ·  HN discussion

🔁 Agents that rewrite their own harness

Picture the part of agent work nobody likes: you tune the system prompt, the tool descriptions and the memory rules by hand, run the agent, squint at transcripts, and edit again. A group led by Ruiyao Xu, with Chen-Yu Lee and others, published code under google-research that tries to do that loop automatically, on tasks it has never seen, without a ground-truth checker.

The paper calls it recursive harnessing. Task-level recursion runs a Propose, Probe, Compose step: a harness editor proposes edits, candidates are tried on the same task, and a composer builds a refined harness from the contrast between runs. A second level keeps a persistent improvement harness that learns from each task. It edits eight components: system prompt, guardrails, memory, tool guidance, subagents, skills, scripts and plugins. The authors report that one step improves GLM-5 and Gemini 3.5 Flash on three professional benchmarks, by up to 11.5 points on WorkBuddy-Bench, 4.5 on JobBench and 11.8 percentage points on LAB partial pass rate. Those are the authors’ numbers on their chosen benchmarks. The repo supports OpenCode and Codex as executors.

The repository  ·  The paper (arXiv)

🧾 Do the labs inflate their own scores?

Someone got tired of the same question the MiniMax thread asked. They collected every case where a lab published a score and someone independent, a leaderboard, a paper or another lab, measured the same model on the same test, and put the result on one page with a dot for each pair. The site is whatmodel.sh. It is one page and the method is printed on it.

By its own count: 611 pairs, 180 models, 103 tests. Lab scores are 6.4% higher on average, with a median gap of 1.2% and a 95% interval of 3.8% to 9.5% for the mean. When the numbers differ, the lab’s is higher 63% of the time: 375 lab higher, 216 independent higher, 20 ties. The gap depends on the test. Fixed-format tests like MMLU, WinoGrande, ARC and HumanEval barely move. Tests where the setup is part of the result, like the agent harness on SWE-bench Verified or the sampling budget on AIME, show the biggest gaps. We did not recompute any of this.

whatmodel.sh/claims

😬 The Meta engineer and the black box

A user posting as rzk quoted a LinkedIn post from a Meta engineer: "I never felt this uncomfortable doing my job." The code feels like a black box, checking every line would take forever, and the engineer asks how much understanding is enough. The Ask HN got 298 points. It is not a launch, which is why it is here.

Three answers stood out. bitwizeshift: do not let the model design, let it implement a design you drew, so you know how the pieces fit even if you did not write each unit. spacedcowboy, who works on an autovectorising compiler, has the model write a huge pile of tests and runs a release gate that takes hours across five architectures. edwin0 named the gap: when the AI writes both the code and the tests, they can agree with each other and miss the same thing. These are commenters’ practices and claims, not studies.

The Ask HN thread

CERN and Mistral sign a statement of intent. A strategic partnership to apply AI to science. HN had it at 304 points. A statement of intent, so no product yet. CERN →

Laminar, an RL-trained model that finds errors in agent traces. Launch HN, YC S24. We read the headline and the thread title only; unverified beyond that. HN thread →

Plex vulnerabilities exploited to drop a cryptominer. Reported as exploited in the wild. HN had it at 148 points. If you run Plex, check your version. HN thread →