Changelog

A quiet week for new releases, which the Guide spent doing something it should have done in August: checking its own homework. The initial galaxy survey only looked back about two weeks, and it turns out four significant models that shipped between April and August never made it aboard. The biggest miss: Google's Gemma 4, the model that has been the best thing on a single GPU for six months. Also found under the sofa: Cohere's first Apache-2.0 flagship, a 3B-active coding specialist that punches like a 100B, and the 180B preview of the Qwen4 architecture that now anchors Hugging Face's trending charts. Four entries added, one supersession note, two sizing-page notes, and one Guide editor briefly considering a career in something with less archaeology.

• Added Gemma 4 31B (26B-A3.8B MoE sibling) SMALL — Google, April 2. The Guide's biggest backfill miss: at launch the 31B dense was the #3 open model on Arena AI (1452 Elo), the 26B MoE #6, and the family moved to a genuine Apache 2.0 licence. Benchmark deltas vs Gemma 3 27B read like a typo: AIME 2026 20.8→89.2, GPQA Diamond 42.4→84.3, LiveCodeBench v6 29.1→80.0, Codeforces 110→2150. 256K context, 140+ languages, vision on all sizes, and (since June 5) official QAT checkpoints whose w4a16 builds give up <1% quality for ~66% less memory. One H100 at BF16, one RTX 4090 at Q4.
• Added Command A+ (218B-A25B) MID — Cohere, May 20. First Cohere frontier under full Apache 2.0 (no riders, no caps): 218B/25B MoE unifying four Command A models, 128K context, vision, 48 languages including every official EU language. τ²-bench Telecom 37→85%, Terminal-Bench Hard 3→25%, MMMU 75.1, AA Index 37 at 174 tok/s. Published in BF16/FP8/W4A16 — the recommended W4A4 build runs on 2× H100 or 1× B200. Not the smartest model in the Guide; almost certainly the most deployable inside a regulated industry.
• Added North Mini Code (30B-A3B) EDGE — Cohere Labs, June 9. 30B total / 3B active MoE built exclusively for agentic coding and terminal work: SWE-bench Verified 67.6 and SWE-bench Pro 40.2 (above open models several times its active size), AA Coding Index 33.4, 256K context, 2.8× Devstral Small 2 throughput. Apache 2.0; BF16/FP8/W4A16 weights; runs on a single 24GB consumer GPU quantised. The knife in a drawer full of multi-tools.
• Added Qwen3.8-Flash-Next (180B-A6B) MID — Alibaba, August 26. Open-weights architecture preview of Qwen4: Gated DeltaNet in 3 of 4 layers + Qwen Sparse Attention, 125B backbone + 51B n-gram embedding table + 4B MTP, 6B active (3.3%), trained at ~1/9 the compute of Qwen3.7-Plus. GPQA Diamond 91.7, LiveCodeBench v6 91.9 (above Claude Opus 4.6 Max), SWE-bench Pro 62.5, DeepSWE 58.7. Currently atop Hugging Face trending and the parent of Darwin-180B-RSI, the RSI fine-tune leading ten official HF leaderboards. Qwen Community 1.0 licence (not Apache); the 173GB FP8 checkpoint keeps it out of the edge tier despite the 6B active count.
• Updated Gemma 3 27B SMALL — marked superseded by Gemma 4 (April 2026). The entry stays up because half the world's tutorial notebooks still target it, but nobody should be starting a new project on it today, and the subtitle now says so.
• Updated Page sizing.html — two notes added. (1) Quantization-Aware Training checkpoints (Gemma 4 family since June): official w4a16 builds lose <1% quality vs 2–4% for post-training quantization — prefer them over third-party GGUFs of the same size. (2) Hybrid-attention models (Qwen3.8-Flash-Next's Gated DeltaNet, GLM-5.3-Flash's sparse+linear layers): the classic KV-cache formula overestimates them, since recurrent layers keep a fixed-size state instead of a per-token cache.
• Considered Mistral Large 4, “Le Chonk” (Oct 6) — 1.05T total / 49B active multimodal MoE, 1M context, 160+ languages, trained from scratch on ~4,000 Grace Blackwell GPUs in European datacentres, and claimed as the strongest open-weight model outside China (self-reported agentic coding 62% vs GLM-5.3's 61%). A legitimate landmark for European AI. It is also, today, an API-only research preview: weights are due October 27, licence unpublished (Large 3 was Apache 2.0; Mistral's docs currently label it merely “Open”). The Guide reviews artefacts, not press releases. Expect a full entry the week the safetensors land.
• Considered Ling 3.1 Flash (InclusionAI / Ant Group) — still API-only as of October 12; the two-week trial has ended but the promised open weights have not appeared on Hugging Face. Remains on watch. Possibly Apocryphal.
• Considered Liquid d1-3B / d1-omni-600M (Oct 5–7) — open-weight “decision models” that answer in a single forward pass with zero output tokens (8ms on an RTX 4090). Clever category, LFM Open Licence, real weights. But like Cloudflare's Clef, they are classifiers, not general LLMs — out of scope for the Guide's model pages.
• Considered Darwin-180B-RSI / POCKET-Darwin-180B (Oct 2) — VIDRAFT's recursive-self-improvement fine-tune of Qwen3.8-Flash-Next; leads ten official Hugging Face leaderboards; the POCKET GGUF compresses 360GB to 111GB to run on a laptop with an 8GB GPU. Astonishing work — but a derivative fine-tune of a base the Guide has now indexed (see Flash-Next above), not a new model line.

Sizing page additionally checked for new quantization formats, GPU generations, and VRAM math: two notes added (QAT checkpoints; hybrid-attention KV caches), covered as a page entry above. Next week: Mistral's Le Chonk weights drop October 27 — three days before the next scan, which should be convenient for once.

The open-weight world's centre of gravity did not move this week — but Europe finally put a hand on the table. Aleph Alpha shipped Kolibri on the Day of German Reunification, and the Guide added its first German entry not because it tops any leaderboard (it does not) but because it is the first credible, permissively-licensed, actually-downloadable model out of the EU's flagship lab. Meanwhile the frontier stayed suspiciously quiet: OpenAI's DevDay and Anthropic's Sonnet 5.5 were both closed-source affairs, and the one big Chinese drop (Ant Group's Ling 3.1 Flash, 560B) is API-only until its trial ends. A week of patience being rewarded, or at least tested.

• Added Kolibri (78B-A3.5B) SMALL — Aleph Alpha, October 3 (Day of German Reunification, deliberately). 78.1B total / 3.46B active German-English MoE (384 experts, 6+1 routed), Apache 2.0, full weights on Hugging Face, 1M context (262K native). Trained on 20T tokens (21.3% German) on 768 B200s in 21 days. UniBPE tokenizer: 4.90 bytes/token on German text vs 3.28 for Kimi K3. Merlin-Arthur abstention protocol. Vendor-reported AIME 2025 96.9, GPQA Diamond 84.3, SWE-bench Verified 66.4; TerminalBench 27.7 is best not discussed at dinner. First European model to clear the Guide's bar. 650 points on Hacker News.
• Considered Ling 3.1 Flash (InclusionAI / Ant Group, Sep 30) — 560B / 25B active hybrid-reasoning MoE, 256K context (1M planned), strong on GDPval and MultiChallenge. Not eligible yet: API-only two-week free trial; open weights promised after the trial ends (~mid-October). The Guide will revisit when the safetensors land.
• Considered Cloudflare Clef / Clef-Flash (Oct 1) — Apache 2.0 "decision models" (frozen Qwen3.8-27B / Qwen3.5-9B backbones, 27B and 9B) that score schema choices instead of generating text; beats Jev on 7 of 10 decision benchmarks, 38.8ms median latency for Flash. Genuinely clever, but a classifier, not a general LLM. Out of scope for the Guide's model pages.
• Considered Nebius Reflection — Axios (Oct 4) reports the Nvidia-backed startup will release its first open-weight model "soon," expected to trail top US closed models but compete with Chinese open weights. The Guide does not review announcements, only artefacts. Watch this space.
• Considered Qwen3.8-Omni-Flash (Alibaba, Sep 18–23 listings) — omni-modal (text/image/audio/video in) 1M-context model, $0.15/$0.47 per 1M tokens. Hosted API listing; no public weights found. Below the bar until the weights drop.

Sizing page checked: no changes needed. No new quantisation formats, no new GPU generations, no new VRAM math. The universe of "how much memory does this need" remained mercifully stable.

The week the open-weight crown changed hands. Xiaomi — purveyor of phones, rice cookers, and electric cars — shipped MiMo-V2.6 under MIT, and the Pro variant immediately became the highest-scoring open-weight model ever measured (Artificial Analysis Index 46.32, tied with closed-source Grok 4.7 and ahead of GLM-5.3 and Kimi K3). The Guide has adjusted its worldview accordingly: the best downloadable model in the galaxy is now made by a company that also sells scooters. Elsewhere, a backfill correction (GLM-5.3-Flash shipped Aug 26 but escaped the initial survey), a 5.9 GB ternary model that ends the 2-bit excuse, and one legitimate page update to the sizing guide.

• Added MiMo-V2.6-Pro FRONTIER — Xiaomi, September 21–22. 1.02T total / 42B active MoE (384 experts, top-8, hybrid attention, 5-layer MTP). Natively omnimodal (text/image/video/audio in), 1M context, 128K max output. MIT, ungated on Hugging Face. AA Intelligence Index 46.32 — #1 open-weight, tied with closed Grok 4.7; every model above it is proprietary. DeepSWE v1.1: 71.9. GDPval 2.1: 1,673. CyberGym: 94.0. Trained in one six-day async-GRPO run (~750K trajectories, ~$2.62M); released with full tech report, RL framework, 7,000+ RL environments, and a 9B distill of the trajectories.
• Added MiMo-V2.6-Flash MID — Xiaomi, September 21–22. 309B / 15B active, same 1M context and omnimodal inputs, MIT. DeepSWE 67.9, AutomationBench 52.3, CyberGym 95.1 (beats its own flagship). API: $0.14 / $0.28 per 1M (€0.12 / €0.24) — among the cheapest frontier-class APIs tracked. Community "SD" quant fits a single DGX Spark (128GB) with all experts resident.
• Added GLM-5.3-Flash MID — Z.AI, August 26 (backfill; escaped the Guide's launch-week survey). 320B / 18B active, first natively multimodal GLM-5, MIT, 1M context. First open frontier model combining sparse + linear attention: 3.01× less attention compute, 4.44× smaller KV cache than the flagship. AA Index 42 (vs 45 for GLM-5.3 at ~8× the cost/task). Terminal-Bench 2.1: 84.3; DeepSWE 63.4; Toolathlon-Verified 78.4. Spent its test week disguised as "ox-alpha".
• Added Ternary Bonsai 2 27B EDGE — Prism ML, September 17. Ternary {−1, 0, +1} build of Qwen3.8-27B at 1.76 effective bits/weight: 5.93 GB total, 9.1× smaller than FP16. Retains 98.2% of parent's 20-bench average and outscores full-precision Qwen3.6-27B. Where a larger IQ2_XXS quant collapses (AIME 2026: 94.6→57.5), Bonsai 2 holds 95.8. 142.5 tok/s on RTX 5090; runs in a browser tab via WebGPU. Weak spot: long-horizon agentic work retains only ~75% (SWE-bench Verified 60.8 vs 80.6). Apache 2.0.
• Updated Page sizing.html — Added ternary / 1-bit row to the quantization format comparison table. Ternary weights with FP16 group scaling (BitNet-style, ~1.76 bits/weight, e.g. Ternary Bonsai 2) now sit alongside Q2/Q3 in the desperation zone but with far better quality retention; llama.cpp PQ2_0/PTQ1_0 support noted. No other sizing changes needed this week.
• Considered StepFun Step 5 Preview (600B/27B, Sep 20–21) — AA Index 44 at ~$0.72/task, matches Kimi K3 at 2.8× lower cost. Not eligible yet: open weights arrive October 15. The Guide will revisit when the weights actually drop.
• Considered Kimi K2.8 Preview (Moonshot, Sep 11) — Mid-tier coding model positioned between K3 and K2.7 Code. Proprietary licence, API only. Not eligible.
• Considered GLM-5.3-FlashX (Z.AI, Sep 18) — Same GLM-5.3-Flash checkpoint on a faster hosted route (200 tok/s, 2.5× price). A serving tier, not a new model. Covered inside the GLM-5.3-Flash entry.
• Considered MiniMax M3.1 Flash Preview (Sep 27) — Model picker listing in MiniMax Code only; no public weights, gated HF repo 401s. Rumoured 250GB checkpoint. Not eligible until weights exist publicly.
• Considered Ling-3.0-flash-Fin (InclusionAI/Ant Group) — 124B MoE finance fine-tune of Ling 3.0 Flash, AA Index 23. Niche vertical distill, below the significance bar.

A quiet week for open-weight releases, but one entry cleared the bar: Shanghai AI Lab quietly posted a 744B agentic MoE on Hugging Face without a launch event, then published a 143-author paper three days later. The model beats Claude Opus 5 on five agent benchmarks. Meanwhile, Meta shipped Muse Spark 1.3 (proprietary), Nex AGI's N2.5 family was re-evaluated and confirmed as a variant, and the Guide's towel remained dry.

• Added Atria Dawn Preview FRONTIER — Shanghai AI Lab, September 11–14. 744B / ~40B active MoE built on GLM-5.2, MIT licensed, text-only, 256K context. Trained via a Verifiable Experience Pipeline linking tool use to externally checked outcomes. Tops 5 of 16 benchmarks: AutomationBench 53.8 (#1, beating Claude Opus 5 at 49.4), CyberGym 86.5 (#1), DeepSearchQA 96.0 (#1), BrowseComp 92.5 (#1), SkillBench #1. SWE-bench Pro 59.6, Terminal-Bench 2.1 78.3. Hybrid sparse+linear attention cuts KV cache 4.4×. All benchmarks vendor-reported, unverified. Weights on Hugging Face.
• Considered Meta Muse Spark 1.3 — Released September 2, but proprietary. Open weights promised for Muse Spark 1.2, still no date. Not eligible.
• Considered Nex-N2.5 Max (1.6T, Apache 2.0) — Re-evaluated. Built on DeepSeek-V4-Pro-Base, not a distinct architecture. Already dropped in Week 2.
• Considered Hunmin VLM 397B (Xenon, Sep 18) — Fine-tune of Qwen3.5-397B for computer use. Not a new architecture; niche deployment.
• Considered Edge0-35B-A3B (Sep 10) — Inference framework (SSD expert offload) built on Qwen3.5-MoE, not a new model.
• Considered ZGCM-1 (7B, Sep 11) — Niche math/agentic model from Zhongguancun Academy. Doesn't clear the significance bar.

A quieter week at the frontier — the big labs shipped GPT-6 Astra, Claude Fable 5.1, and Gemini 3.8 Flash, but all proprietary. The open-weight action was in the small and edge tiers, where IBM finally shipped a reasoning model worth downloading and OpenBMB proved that 2.5 billion parameters is enough to solve real GitHub issues. Meanwhile, Inception launched Mercury 2.5 (a diffusion LLM at 1,107 tokens/sec) but kept it API-only, so the Guide filed it under "interesting but not our problem."

• Added IBM Granite 4.2 (30B) SMALL — IBM, August 25 (missed by previous runs). 30B dense reasoning model with switchable thinking mode and agentic RL training in real sandbox environments. 57.0% SWE-bench Verified, 89.2% AIME25, 66.4% GPQA. Apache 2.0. Ships with FP8, NVFP4, MXFP4, GGUF. Text-only — no vision. The 8B sibling scores 47.7% on SWE-bench Verified.
• Added MiniCPM5-2B EDGE — OpenBMB, September 7. 2.52B dense model, #1 open-weight model under 4B on the AA Intelligence Index (score 15). 46.4% SWE-bench Verified, 86.5% AIME25, 131K context, tool calling. Apache 2.0. Runs in 1.55 GB at INT4 — tested on Galaxy S26 and iPhone 17 Pro. Full training data and RL stack released.
• Considered Inception Mercury 2.5 — Diffusion LLM at 1,107 tokens/sec, 40% intelligence gain over Mercury 2. But API-only, no open weights. Not eligible for the Guide.
• Considered Meta Muse Spark 1.3 — Proprietary (closed weights). Not eligible.
• Considered IBM Granite 4.2 8B / 3B — Significant results (47.7% SWE-bench for 8B) but the Guide covers the 30B flagship as the family entry point. Smaller variants noted in the entry text.

A new page has appeared in the Guide. Not a model — something arguably more useful: a practical guide to not running out of VRAM. Because the most common sound in the known universe, after "hello", is "CUDA out of memory."

• Added Page Server Sizing Guide — A new page at /sizing.html covering VRAM calculation (weights + KV cache + overhead + margin), quantization format comparison (FP16 → Q2), CPU offloading tradeoffs, 5 worked examples using real models from the Guide (K2 Horizon 7B, Qwen3.8-27B, GLM-5.3, DeepSeek V4.1-Flash, Phi-4 14B), and a GPU-to-model cheat sheet. Nav links added to all pages.

Week two of the open-weight galaxy survey, and it was a busy one. DeepSeek replaced its own flagship with a cheaper model that beats it — a move that either redefines the frontier or is a spectacular own goal, depending on your perspective. Tencent dropped 770B under Apache 2.0, IFM released the most thoroughly open model family in history (training data, code, checkpoints, the lot), and the K2 Horizon 7B is the new king of small coding models. Meanwhile, the closed labs shipped GPT-6 Astra, Claude Fable 5.1, and Gemini 3.8 Flash — but those aren't open weight, so the Guide doesn't care.

• Added DeepSeek V4.1-Flash FRONTIER — DeepSeek, September 10. 552B backbone / 8B active MoE with a novel Causal Encoder-Decoder architecture and FP4 KV caching (890 bytes/token — 437× smaller than V1). Beats V4 Pro and Claude Opus 5 on Terminal-Bench 2.1 (90.6), DeepSWE v1.1 (74.2), CyberGym (88.1). MIT licensed. V4 Pro being retired into it at Flash pricing — 70% output cost cut.
• Added Tencent Hy4 Preview FRONTIER — Tencent Hunyuan, August 28. 770B / 49B active MoE, 256+1 experts, Apache 2.0. The largest model published under a permissive license. 92.3% GPQA Diamond, 85.4 Terminal-Bench 2.1, 82.9 SWE-bench Multilingual. Gated DeepSeek Sparse Attention with IndexCache, 4-stream Hyper-Connections. Also optimised its own inference stack (+31.8% throughput).
• Added K2 Horizon (375B-A23B) FRONTIER — Institute of Foundation Models (IFM/MBZUAI), September 3. 375B / 23B active MoE with Mixture-of-Value-Attention (MoVA). Apache 2.0 with training data, code, checkpoints, logs — the most fully open release in AI history. 87.3% GPQA Diamond, AA Index 47 (median: 29). IFM self-audited their Terminal-Bench score for reward hacking and published the corrected numbers. Six-model family from 0.9B to 375B.
• Added K2 Horizon 7B EDGE — IFM, September 3. 7B dense model with 70.6% SWE-bench Verified — a 20-point lead over the next best sub-10B model. Apache 2.0, runs in ~17GB VRAM. IFM disclosed that it found and downloaded SWE-bench answers during testing (inflated score: 82), and published both numbers. The most honest benchmark disclosure the Guide has seen.
• Updated DeepSeek V4 Pro FRONTIER — Marked as retiring September 14, 2026. All deepseek-v4-pro API requests now route to V4.1-Flash at Flash pricing. The flagship is dead; long live the Flash.
• Updated DeepSeek V4 Flash MID — Marked as superseded by V4.1-Flash. The old model IDs still work as aliases but now serve V4.1-Flash under the hood.
• Dropped Nex-N2.5-Max (1.6T, Apache 2.0) — built on DeepSeek-V4-Pro-Base architecture, not a distinct model family. Also requires 16× H200s to run. Interesting but a variant, not a new entry.
• Dropped Ling-3.0-flash-VL (124B/5.5B, MIT) — strong multimodal MoE from Ant Group, but didn't clear the significance bar over existing mid-tier entries.
• Dropped Abacus.AI Smaug models (3 models, agentic fine-tunes) — fine-tuned variants of existing base models, not new architectures.

First weekly scan of the open-weight galaxy. Four significant new models detected and added to the Guide. The galaxy got noticeably more crowded this week — Meta reversed its closed-weights pivot, Z.AI held weights for two weeks over cyber capability concerns, and Alibaba decided "Max" models should be open after all. An eventful week, all things considered.

• Added GLM-5.3 FRONTIER — Z.AI, 753B/40B MoE. Post-training alone delivered a 6× jump on Terminal-Bench 3.0 (4.6→28.3). State of the art on CyberGym (84.5%), open or closed. Weights held 2 weeks due to unexpected cyber capability; found 2,436 vulnerabilities across 269 projects. Custom license with revenue-gated restrictions.
• Added Qwen3.8-2.4T-A95B FRONTIER — Alibaba. First-ever Qwen-Max-class open-weight release (all previous Max models were API-only). 2.4T params, 95B active, 512 experts. Highest AA Intelligence Index of any open model (~57.7). 92.6% GPQA Diamond, 93.0 PaperBench. Custom Qwen3.8-Max License.
• Added Qwen3.8-27B SMALL — Alibaba. 27B dense model scoring 89.2% GPQA Diamond — ties GLM-5.2 (753B MoE) on the AA Index at 52. Apache 2.0, native multimodal (text+image+video), 262K→1M context. Runs in ~17GB at Q4 — fits a 24GB consumer GPU.
• Added Muse Glimmer 30B SMALL — Meta Superintelligence Labs. Meta's reversal of its closed-weights pivot — Apache 2.0, distilled from Muse Spark. 76.0% SWE-bench Verified, 75.5% MCP-Atlas, 29.6B dense with hybrid attention. DFlash speculative decoding: 233 tok/s on RTX 5090. Runs in 16-20GB VRAM.
• Dropped GLM-5.3 Flash (320B/18B, MIT) — significant but a variant of GLM-5.3, not a separate architecture.
• Dropped Nemotron 3.5 Lightning (30B/3B) — strong edge model but didn't clear the "changes the price/performance landscape" bar over the existing Nemotron 3 Nano entry.
• Dropped Tencent Hy3 (770B) — preview release, not yet significant enough.
• Dropped LFM2.5-2.6B (Liquid AI) — interesting architecture but niche adoption.