ModelsMistral Medium 3.5: One Dense 128B Model for Chat, Code and Reasoning
Mistral merged its chat, reasoning and coding lines into one dense 128B open-weight model. What the merge means, what it costs, and the licence clause to read first.
ReadNotes on models, agents, and what inference costs
Practical writing for developers building with large language models — how they work, how to pick one, and how to keep the bill predictable.
ModelsModelsMistral Nemo Guide: The 12B Model That Outlived Its API
Mistral Nemo is a 12B Apache 2.0 model built with NVIDIA, now retired from Mistral's own API. What it is, why Tekken mattered, and where it still earns a slot.
ReadModelsMistral Saba Guide: A Regional 24B Model and Its Retirement
Mistral Saba was a 24B model tuned for Arabic and South Asian languages, and Mistral has since retired it. What it was for, and how to replace it without a regression.
Read
ModelsMistral Small 3 (24B, 2501): The Latency-First Open Model
Mistral Small 3 is a 24B Apache 2.0 model trained without RL or synthetic data, tuned for speed over depth. Its specs, its 32K limit, and when it is still worth running.
Read
ModelsMistral Small 3.1 24B: Vision and 128K in an Open Model
Mistral Small 3.1 added image input and a 128K window to a 24B Apache 2.0 model. What changed from Small 3, what 3.2 fixed, and the narrow case for staying on 3.1.
Read
ModelsMistral Small 3.2 24B: The Point Release Agents Needed
Mistral Small 3.2 kept the 24B size of 3.1 but cut runaway repetition and hardened function calling. What the numbers show, how to serve it, and what replaces it.
Read
ModelsMistral Small 4: A 119B MoE That Still Calls Itself Small
Mistral Small 4 folds reasoning, vision and agentic coding into one Apache 2.0 MoE with 6B active parameters. What the name hides, what it costs, and how to use it well.
Read
ModelsMixtral 8x22B Instruct: Should You Still Be Running It?
Mixtral 8x22B was a landmark open MoE in 2024: 141B total, 39B active, Apache 2.0. Mistral retired it in 2025. What it was good at, and what moving off it costs.
ReadModelsMorph V3 Fast: An Apply Model, Not a Chat Model
Morph V3 Fast does one job: merging a lazy code edit into a file at over 10,000 tokens a second. How the apply pattern works, what it costs and when to step up to Large.
ReadModelsMorph V3 Large: When Your Apply Model Should Be the Slow One
Morph V3 Large trades speed for merge accuracy. The maths of halving a failure rate, what a bad merge really costs, and a rule for which edits deserve the slower model.
ReadModelsMuse Glimmer 30B: Meta's Apache-Licensed Local Agent Model
Meta's dense 30B multimodal model ships under Apache 2.0 with a speculative drafter. What fits on a 24GB card, where it beats Qwen 3.6 27B and where it trails.
ReadModelsMuse Spark 1.1 Guide: Meta's First API Model, Two Versions On
Muse Spark 1.1 opened the Meta Model API in July 2026 and was superseded within a month. What it does well, where it trails, and when calling it still makes sense.
ReadShowing 205–216 of 731 articles