The model release log

Notable model releases and technique shifts, quarter by quarter. Every entry is dated and carries a primary source.

Q3 2026 (July – September)

The efficiency race sharpened: DeepSeek shipped a 552B-parameter MoE that activates only 8B parameters during prefill, and Qwen pushed a natively multimodal 27B whose context extends to a million tokens. Serving news mattered as much as model news — OpenAI’s frontier flagship went generally available on a rival’s cloud — and Mistral put an open-weight, policy-adaptive safety classifier into the guardrail conversation.

  • 2026-09-14 · serving — OpenAI GPT-6 Astra generally available on Amazon Bedrock. AWS announced that GPT-6 Astra — OpenAI’s most capable model to date — is now generally available on Amazon Bedrock, supporting up to 1 million input tokens and accessible through Bedrock APIs or by pointing ChatGPT Work and Codex at Bedrock. So what: Frontier models and serving platforms are decoupling: the same closed model can now be procured through a rival’s cloud, so provider choice is a serving decision, not just a model decision. (source)
  • 2026-09-10 · model — DeepSeek-V4.1-Flash: 552B MoE with a causal encoder–decoder split. DeepSeek released V4.1-Flash, a 552B-parameter multimodal MoE with a new Causal Encoder–Decoder architecture that activates only 8B parameters per token during prefill (16B during decode) and cuts KV-cache HBM needs to a quarter of the previous generation. It replaces V4-Flash in the API, with V4-Pro traffic routing to it from September 14, 2026. So what: Asymmetric prefill/decode activation is MoE economics taken to its logical end — total parameters, active parameters, and cache footprint are now three separate levers you should read off every model card. (source)
  • 2026-08-13 · model — Qwen3.8: open multimodal family with million-token extended context. Alibaba’s Qwen published the Qwen3.8 collection (the card’s citation dates it to August 2026), led on the small end by Qwen3.8-27B: Apache 2.0, native image and video understanding, a 262,144-token native context extensible to 1,000,000 via YaRN, and thinking mode on by default with a tunable reasoning_effort. So what: Thinking-on-by-default with an effort dial is becoming the open-weight norm too — budget for reasoning tokens even when you self-host, or turn the mode off deliberately. (source)
  • 2026-08-04 · safety — Shieldstral: policy-adaptive open-weight safety classifier. Mistral released Shieldstral, a 3B open-weights (Apache 2.0) multimodal safety classifier that takes moderation policies as plain-language yes/no questions at inference time and returns a calibrated safety score from the yes/no token logits, covering prompts, responses, and images on a single 16GB GPU. So what: Guardrails are shipping as swappable models with policies supplied at inference — you can retarget moderation without retraining, which changes how you architect the safety layer around your LLM. (source)

Q2 2026 (April – June)

Open-weight flagships converged on one recipe: multimodal input, long context, tunable reasoning effort. Google moved Gemma 4 to a plain Apache 2.0 license and followed with quantization-aware checkpoints that squeeze its edge model under a gigabyte, while Mistral consolidated its lineup around a single merged flagship.

  • 2026-06-05 · technique — Gemma 4 QAT checkpoints shrink the edge model to about 1GB. Google released Gemma 4 checkpoints optimized with quantization-aware training: a mobile-specialized format with 2-bit compression of token-generation layers plus embedding and KV-cache optimizations brings Gemma 4 E2B to roughly 1GB (under 1GB text-only), with support across llama.cpp, Ollama, vLLM, SGLang, MLX, and LiteRT-LM. So what: QAT simulates quantization during training instead of applying it afterward, which is why these checkpoints hold quality where naive post-training quantization degrades — the difference matters when you deploy on a phone or a consumer GPU. (source)
  • 2026-05-22 · model — Mistral Medium 3.5: one merged flagship instead of a model zoo. Mistral’s dated news feed introduced Medium 3.5 on May 22, 2026 as the model powering remote agents in Vibe; its model card describes “our first flagship merged model” — a dense 128B with 256k context under a Modified MIT license — replacing Mistral Medium 3.1, Magistral, and Devstral 2, with per-request reasoning effort (none or high). So what: Vendors are collapsing separate chat, reasoning, and coding models into one flagship with a reasoning switch — fewer models to choose between, but license fine print (Modified MIT with revenue exceptions) still decides who can use it. (source)
  • 2026-04-02 · model — Gemma 4 launches under Apache 2.0. Google introduced Gemma 4 in four sizes — E2B and E4B for edge devices, a 26B MoE activating 3.8B parameters at inference, and a 31B dense model — with vision and audio input, 128K–256K contexts, 140+ languages, and a switch from the Gemma Terms of Use to a commercially permissive Apache 2.0 license. So what: A first-party open family dropping its custom license for Apache 2.0 removes the redistribution and fine-tuning caveats that made legal review of “open” models slow — check the license page, not the marketing label. (source)

Q3 2025 (July – September)

OpenAI returned to open weights with the Apache-2.0 gpt-oss pair, published alongside a formal model card paper. The quarter’s quieter theme was trustworthy-by-construction: a token-streaming guardrail family from Qwen and the first 1B-parameter model pre-trained from scratch under differential privacy.

  • 2025-09-23 · safety — Qwen3Guard: guardrails that classify the token stream in real time. Qwen released its first guardrail family in 0.6B, 4B, and 8B sizes: Qwen3Guard-Gen for offline annotation and RL safety rewards, and Qwen3Guard-Stream, which scores tokens as they are generated via two lightweight classification heads, using a three-tier Safe/Controversial/Unsafe scheme across 119 languages. So what: Moderating a stream token-by-token instead of after the full response lands is what lets you cut off unsafe output mid-generation — a pattern worth knowing before you build your own output filter. (source)
  • 2025-09-12 · technique — VaultGemma: 1B model pre-trained entirely with differential privacy. Google Research released VaultGemma, described as the largest (1B-parameter) open model trained from scratch with differential privacy — DP-SGD guided by new DP scaling laws — carrying a sequence-level guarantee of ε ≤ 2.0, δ ≤ 1.1e-10, with no detectable memorization of training data. So what: It quantifies the current privacy tax: utility comparable to non-private models from about five years earlier — a concrete anchor for any conversation about training-data memorization and privacy guarantees. (source)
  • 2025-08-08 · model — gpt-oss: OpenAI returns to open weights. OpenAI published the model card paper for gpt-oss-120b (117B parameters, 5.1B active) and gpt-oss-20b (21B, 3.6B active) — Apache 2.0 mixture-of-experts reasoning models trained with large-scale distillation and RL, shipped with MXFP4-quantized MoE weights, adjustable reasoning effort, and a mandatory harmony response format. So what: A frontier lab’s reasoning stack in downloadable form: the 120b fits one 80GB GPU, so the open-versus-API decision now includes genuinely strong reasoning models you can host yourself. (source)

Q2 2025 (April – June)

Meta’s Llama 4 brought mixture-of-experts and native multimodality to the open-weight mainstream — under a community license with real strings attached. Anthropic’s Claude 4 generation leaned into long-running agentic coding, and Google announced its first TPU designed specifically for inference, a signal that serving economics now drive silicon design.

  • 2025-05-22 · model — Claude Opus 4 and Sonnet 4: extended thinking meets tool use. Anthropic introduced Claude Opus 4 (billed as the world’s best coding model, 72.5% on SWE-bench) and Claude Sonnet 4, with a beta letting both models alternate between reasoning and tools like web search during extended thinking, plus parallel tool use and hours-long agent sessions. So what: Interleaving thinking with tool calls is the pattern behind modern agent loops — the same reason your agent’s token bill is dominated by reasoning and tool-result tokens, not the final answer. (source)
  • 2025-04-09 · hardware — Ironwood: Google’s first TPU designed specifically for inference. Google announced Ironwood, its seventh-generation TPU and the first built specifically for inference: 192GB of HBM per chip at 7.37 TB/s, scaling to 9,216 liquid-cooled chips per pod (42.5 exaflops), with double Trillium’s power efficiency, available to Cloud customers later in 2025. So what: When an accelerator generation is designed around serving rather than training, it confirms where the cost now lives — inference throughput and memory bandwidth are the numbers that set your per-token price. (source)
  • 2025-04-05 · model — Llama 4: open-weight MoE goes natively multimodal. Meta announced Llama 4 Scout (17B active parameters, 16 experts, 109B total) and Maverick (17B active, 128 experts, 400B total) — its first natively multimodal, mixture-of-experts Llama models, with Scout claiming a 10M-token context — under the Llama 4 Community License effective the same day, which requires “Built with Llama” attribution and a separate license above 700 million monthly active users. So what: The flagship example of why “open-weight” is not one thing: architecture headlines (MoE, 10M context) come bundled with license terms you must read before shipping a derivative. (source)

Q1 2025 (January – March)

DeepSeek-R1 showed that reasoning can be incentivized with pure reinforcement learning — and published the recipe, resetting expectations for open labs. Around it, Mistral argued that a fast 24B can replace closed mid-tier APIs, and Anthropic demonstrated classifier-based jailbreak defenses that survived thousands of hours of red-teaming.

  • 2025-02-03 · safety — Constitutional Classifiers: jailbreak defense that held under red-teaming. Anthropic published Constitutional Classifiers — input and output classifiers trained on synthetic data generated from a constitution of allowed and disallowed content. In automated tests the jailbreak success rate on Claude 3.5 Sonnet fell from 86% to 4.4%, with overrefusals up only 0.38% and a 23.7% compute overhead; 3,000+ hours of human red-teaming against the prototype found no universal jailbreak. So what: Defense-in-depth for LLMs looks like this in practice: measured trade-offs between attack success, overrefusal, and serving cost — the three numbers to demand from any guardrail you evaluate. (source)
  • 2025-01-30 · model — Mistral Small 3: a 24B built to replace closed mid-tier APIs. Mistral released Small 3, a latency-optimized 24B-parameter model under Apache 2.0 (base and instruct checkpoints), claiming over 81% on MMLU at 150 tokens/s and positioning it as “an excellent open replacement” for closed models like GPT-4o-mini — notably trained without RL or synthetic data. So what: The candid “80% of generative AI tasks” framing is a useful sizing heuristic: most workloads need strong instruction-following at low latency, not a frontier model — and a quantized 24B runs on a single RTX 4090. (source)
  • 2025-01-22 · technique — DeepSeek-R1: reasoning incentivized by pure reinforcement learning. DeepSeek-AI submitted the DeepSeek-R1 paper to arXiv (v1 on January 22, 2025), claiming that reasoning ability can be developed through pure RL without human-labeled reasoning trajectories, with emergent self-reflection, verification, and dynamic strategy adaptation; the listing records later journal publication in Nature 645. So what: This is the openly documented recipe behind the reasoning-model wave — when a model card mentions RL-trained “thinking”, this paper is the lineage to understand first. (source)