ModelsDeepSeek V3 0324: Same Architecture, Better Post-Training
DeepSeek V3 0324 kept the V3 architecture and changed the training on top: large reasoning gains, an MIT licence, and a temperature mapping that still catches people out.
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.
Models
ModelsDeepSeek V3.1 Terminus: Reading a Patch Release Honestly
Terminus is a September 2025 fix-up of DeepSeek V3.1: fewer mixed-language outputs, stronger agent scores and a few quiet regressions. What changed and whether to switch.
Read
ModelsDeepSeek V3.1: One Model With Thinking and Non-Thinking Modes
DeepSeek V3.1 merged the chat and reasoner lines into one hybrid model. How the two modes are switched, what each one scored, and which mode handles which tool calls.
ReadModelsDeepSeek V3.2 Guide: Reasoning Inside Tool Calls, MIT Weights
DeepSeek V3.2 was the first DeepSeek model to think while calling tools. What that changed for agents, what Speciale was, and whether V3.2 still earns a place in 2026.
ReadModelsDeepSeek V3.2-Exp: The Sparse Attention Experiment, Explained
V3.2-Exp tested DeepSeek Sparse Attention on a 685B MoE and cut API prices by more than half. What DSA changed, what the scores showed, and why V3.2 replaced it.
ReadModelsDeepSeek V4 Flash Vision Exp: A Vision Experiment, Now Retired
DeepSeek added image input to V4 Flash as an experiment, then retired it on its API three weeks later. What it was, how to read its scores, and what to do if you call it.
Read
ModelsDeepSeek V4.1 Flash: Cheap Prefill for Agents That Read a Lot
V4.1 Flash splits a 552B MoE into an encoder and a decoder, with 8B active on input and 16B on output. Why that suits agents, what it beat, and what it retired.
ReadModelsDevstral 2 Guide: Mistral's 123B Dense Coding Agent
Devstral 2 is a 123B dense model at 72.2% on SWE-bench Verified, with a modified MIT licence that caps company revenue. Where it fits, and when Small 2 is the better pick.
ReadModelsDolphin Mistral 24B Venice Edition: What Uncensored Really Means
Venice Uncensored is a Mistral Small 24B fine-tune that hands alignment to whoever writes the system prompt. What that buys you, what it costs, and when to use it.
Read
ModelsDots3-Note Preview: RedNote's 280B Apache-Licensed Omni Model
Dots3-Note Preview is a 280B MoE with 16B active, a 512K window and Apache 2.0 weights from RedNote's dots studio. What is in it, what is claimed, and how to test it.
ReadModelsEmber-1 Guide: Kimi K3 Quality With Fewer Reasoning Tokens
Fireworks post-trained Kimi K3 to think less and claims the same quality with about 40% fewer tokens. What Ember-1 is, how to read its numbers, and how to test it.
ReadModelsERNIE 4.5 VL 424B A47B: Baidu's Open Vision MoE in 2026
Baidu's largest open vision-language model is a 424B MoE with 47B active and separate text and vision experts, under Apache 2.0. What it is and whether it still earns a slot.
ReadShowing 37–48 of 731 articles