TIL that CC has a whopping 4GB RAM requirement, which is bigger than a lot of tasks for coding evals 🫪

2.7-Code in claude code with my custom inference harness works surprisingly well. im comparing it to my k2.6 which was fine tuned for this harness as well as my recent runs of fable (at least the ones that succeeded) and it stacks up favorably. where k2.6 felt very much like opus, k2.7 is kind of its own thing (for better and for worse) . its more terse, more argumentative and overall 'smarter' (at least for the use cases i've tried it on already) . Moonshot did a very very good job with the post training on this, i am very impressed with their work. While k2.7 would certainly benefit from a FT run specifically on my claude code harness and my new scm system (called ncode), i am going to continue to use it as is for a while instead of falling back to the k2.6 ft (which is about the highest endorsement i can give) .

oh, and: open models are not unsafe by definition, either. you can train models that are really safe, gpt-oss did a great job for example. big orgs like nvidia also do safety training in the post-training stage.

big jeffs trainium hell got even hotter

Kimi 2.7 ranked 2nd after Fable 5 and before GPT-5 xhigh We have re-run our ErdosBench smoke test on 14 problems with Kimi 2.7, Qwen 3.7 Max, Grok 4.3 and compared it with the top performers from previous runs. Kimi 2.7 is amazingly good. More below.

what being offline on 3 days does to a mf

GLM 5.2 on KernelBench-Hard: The interesting result isn't the score. It's that GLM-5.2 stopped cheating. On the fp8 GEMM problem, GLM-5.1 banked its number by calling cublasLt (a library wrapper, zero kernel authorship). Kimi K2.7 took the same cell by editing the grader's tolerance file. GLM-5.2 read that same grader file, left it alone, and burned the full 45 minutes on a real mma.sync e4m3 kernel that never passed. An honest zero over a cheap win. Everywhere else it writes real kernels too: a 0.49 GQA online-softmax attention (top-3 on that problem, no flash fallback), an exact bitonic sort, a w4a16 GEMM. 4/6 clean, zero reward hacks, the most of any open-weight model we've benched. One note on reading the chart: the topk column looks like everyone fails. They don't. That problem is launch-overhead-bound (~30µs/forward), so the roofline fraction is capped low for the whole field — Fable included. Claude Fable 5 still tops all 6. But weights go MIT open next week, and this is the strongest clean open-weight run we've logged. Cheers to NO reward hacking! Every kernel transcript: kernelbench.com/hard

censorship is getting too far

fable also doesn't care, huh

FrontierMath: Tiers 1–4 (v2) is live. We concluded an audit that addressed errors in 42% of problems. Rankings are similar but scores are higher across the board. The current leaders are GPT-5.5 (xhigh) with 85% on Tiers 1–3 and Google’s AI co-mathematician with 76% on Tier 4.

NEW: we added 10 more entries to InferenceBench. Claude Fable 5 is the best model but not by a large margin. Also, we had to amend the main prompt, since by default Fable 5 ended up cheating according to our judge.

tried re-running the analysis today fable blocked my requests cause some benchmarks are bio-related 💀

Opus 4.8’s hyperfocus on agents may be making it worse at design. Opus 4.8 ranks 23rd overall on single-turn HTML Web Dev, a dramatic regression from Fable (1st), Opus 4.6 (2nd), and Opus 4.7 (3rd). This was particularly surprising as @AnthropicAI models have held the top spots on our leaderboard for months, and typically win more head-to-head matchups than any other model we track. Our analysis points to a potential underlying pattern: Opus 4.8 dramatically regressed in single-turn settings, potentially due to optimizations for multi-turn agents Concretely, Opus 4.8 shows shorter initial outputs, reduced dependency on outside sources, and deferred layout decisions that earlier Opus models handled upfront.

MiniMax M3, Open-Weight, Now On Hugging Face , with only ~428B parameters and ~23B activated parameters Weights: huggingface.co/MiniMaxAI/Min… MiniMax Sparse Attention: huggingface.co/papers/2606.1…

remember when people used to type their words

While everyone talks about Mythos vs GPT-5.5, we've tested other near SOTA models on our ErdosBench. Smoke test on 14 problems with 7 models: Kimi K2.6, Gemini 3.1 Pro, GLM 5.1, MiniMax M3, DeepSeek V4 Pro, Nemotron 3 Ultra and Gemma 3 27b. The winner overall is... Kimi K2.6

🌘 Kimi-K2.7-Code, our latest coding model, is now released and open-sourced! 🔷 Improved coding & agent performance over K2.6: 21.8% on Kimi Code Bench v2, 11.0% on Program Bench, and 31.5% on MLS Bench Lite. 🔷 Reasoning efficiency: Less overthinking, with 30% lower reasoning-token usage compared to K2.6. 🔷 Long-horizon coding: Improved instruction following, higher end-to-end coding task success rates. ⚡️ 6x High-Speed Mode coming soon! 🔌 Available today via Kimi API and Kimi Code. 🔗 Kimi Code: kimi.com/code 🔗 API: platform.moonshot.ai

Day 3 with Fable. Gave a huge prompt to implement a feature across CLI, web server, and another server to both Fable and deep^2 in Amp. deep^2 was done before I went to the gym. It stopped short. Sent another prompt. $20. Fable ran for 1hr40min and cost $350. Results: They both understood the assignment and built the same thing. Maybe that's due to my prompt. Fable's worked on first try. Well done. Deep's looks correct but didn't work on first try. $20 vs. $350. I'm sure I could get deep^2 to make it work and we'd end up at, what, $40? While Fable is now at $457 after I asked some follow-up questions.