Content History
Upon request, here's an updated version with Grok 4.5 and Meta's Muse Spark 1.1. Grok 4.5 seems to sit at the Pareto frontier. Good bang for the buck. (Also added harness info).
For agentic coding, one can say: - Unless you need Terra Ultra perf, it's always better to use a Luna model with higher effort setting (same or better performance but cheaper). - Forget everything below Sol High, use Luna with higher effort settings here - Forget Sol Extra High, use Terra Ultra here - The extra cost of Sol Ultra is probably not worth it over Max
I like choices... but now I have: 2x modes (Codex vs. Work mode) 3x GPT-5.6 models (Sol, Terra, Luna) 5x effort levels (Light, Medium, High, Extra High, Ultra) That's 2 x 3 x 5 = 30 possible configurations for a query 🤯 What happened to "Auto" mode?
Have been taking different local open-weight LLMs for a test drive in different harnesses (Qwen-Code, Codex, Claude Code). 30B Mixture-of-Expert models are kind of a nice sweet spot and can solve challenging problems. And they get roughly 40 tok/sec on a Mac or DGX Spark, which is similar to GPT 5.5 in a Pro subscription and totally useable for everyday work. More interesting is also the harness choice! Claude Code seems to be using 2x many tokens as Codex. Gemma 4 E2B is here just for reference to show that the tasks can't be trivially solved by smaller models. Just finishing a longer write-up about this and will share soon (likely tomorrow)!
Just caught up with the recent GLM-5.2 release. The best open-weight model today. Architecture-wise, it's build on the GLM-5 and GLM-5.1 architecture that I covered previously, which means it's reusing the Multi-head Latent Attention (MLA) and DeepSeek Sparse Attention (DSA) mechanisms from DeepSeek V3.2. (I wrote about it here: https://magazine.sebastianraschka.com/p/technical-deepseek) What's new is that they added an IndexShare mechanism. (That's a cross-layer reuse trick for DSA where instead of recomputing the sparse-attention top-k indexer in every layer, GLM-5.2 runs the full indexer only once every four layers and lets the following layers reuse those selected token indices. This keeps the same DSA idea but makes 1M-token inference much cheaper.)
Crazy model! It actually uses the old Qwen2.5-Coder-3B stack and got really great performance with their post-training stack. Need to use it in the next days to see if vibes of VibeCoder actually check out in practice. But impressive first impression! Based on the tech report, some of the important pieces of their post-training stack: 1. High-signal synthetic data (math problems with credible solutions, code with tests) 2. Multiple reasoning paths for each answer 3. Filtering, filtering, filtering 4. 2-stage SFT (start with broad training, then train on hard long-reasoning samples) 5. Use target (pass@k) accuracy over validation loss for checkpoint selection 6. MGPO (MaxEnt-Guided Policy Optimization) for RLVR: basically a GRPO-style RL method with an extra weighting that favors examples that are neither too easy nor too hard for the current policy 7. Single 64k long-context RL (they found that the usual progressive context expansion hurt this model because early truncation damaged long-thinking behavior) 8. Training data order: they do Math RL, then Code RL, then STEM RL in this particular oder which they found helped overall 9. After optimizing for accuracy, they add a stage that rewards shorter correct trajectories; basically making the model more efficient without accuracy degradation
WHAT THE HELL is happening in AI? A 3B parameter model just put up coding benchmark scores in the same league as Claude Opus 4.5. 3 BILLION. The weights are on Hugging Face, anyone can test it. I genuinely don't know if this is a breakthrough or if the benchmarks are broken.
Cool new open-weight model by Cohere: a new lightweight 30B open-weight model for agentic coding tasks. This one builds on Command A+ using the parallel transformer design. Interestingly, even though it's almost half as big, it almost doubles the number of layers. Also, they say that it's been specifically developed for agentic coding, not just coding. I.e., the evaluation is inside a workflow, not just on a single prompt-to-code-answer task. For Terminal-Bench, the model has to use a terminal, inspect the environment, run commands, read outputs, etc. For SWE-Bench the model works on real GitHub-style software issues where it has to understand the repository, find relevant files, make a patch, pass tests, etc. SciCode and LiveCodeBench are more traditional because they mostly test whether the model can produce correct code for a specified problem. Sure, this still requires reasoning, but it's more like “Implement a numerical routine to compute a scientific quantity from given equations and inputs.” which doesn't require any interaction with the environment, existing files, tests, etc. The focus on the agentic code benchmarks is probably why it's far ahead of Gemma 4 on those. Overall, it's pretty competitive although not quite Qwen3.6-level performance.
Turns out Fable 5 is shadowbanning AI researchers 🫤
mythos will be bad ON PURPOSE on ai "frontier llm research" tasks, this is very very sad for the research community also the fact that this is un purpose not visible to the user is crazy
Always back to the basics: LatentMoE was probably inspired by MLA, which was inspired by LoRA, which was inspired by SVD, which was inspired by eigendecomposition.
