Open models are a separate frontier with its own labs, licenses, and release cadence — so this Watch tracks them separately from the Commercial Watch: who is shipping, what the standings say, what the licenses actually allow, and what landed this month. Standings below are dated snapshots from public leaderboards, with sources.
The labs whose weights you can download. The open frontier turns over roughly every six weeks; these are the names at the top of it in October 2026.
Flagship: MiMo-V2.6-Pro. Tops the Artificial Analysis Intelligence Index among open weights (46.32, v4.3.2) under a plain MIT license.
Flagship: GLM-5.3. 44.78 on the Index, positioned for agentic coding; the smaller GLM-5.3-Flash ships MIT. The flagship’s own license carries revenue-gated conditions.
Flagship: Kimi K3. 43.59 on the Index, a long-context, long-horizon coding model; #1 on LMArena’s Frontend Code Arena in July 2026. Bespoke license terms.
Flagship: DeepSeek V4. Under MIT, 1M-token context. V4-Pro posts 80.6% on SWE-bench Verified — the top open score on that board. V4.1-Flash-1017 is the current efficiency tier.
Flagship: Qwen3.8. The widest size range in the field: from a 27B dense model (Apache 2.0) to a 2.4T MoE flagship (bespoke license).
Flagship: Mistral Large 4. "Le Chonk" — 1.05T parameters, 49B active, 1M context — entered public preview Oct 6; weights are due end of October. Medium 3.5 reports 77.6% on SWE-bench Verified (developer’s own figure).
Flagship: Nemotron 3.5. Family (Oct 15–16): Super VL at 120B, and Bolt 74B-A9B, which NVIDIA says fits 24GB of memory at native 2-bit quantization. Open weights, datasets, and recipes.
Flagship: Muse Glimmer 30B. The flagship (Muse Spark) is API-only now. The current open-weight release is Muse Glimmer 30B, Apache 2.0, built for local agents; Llama 4 remains under the Llama Community License.
Flagship: Ornith-1.5. 397B MoE, 35B-A3B, and a dense 9B, MIT-licensed, trained by an end-to-end self-improvement loop (self-generated tasks, self-built scaffolds, RL rollouts). The 9B is this desk’s go-to local model — read the feature.
Flagship: Gemma 4 · gpt-oss-120B. Gemma 4 (Apache 2.0) is the local/edge pick, down to phone-sized checkpoints. gpt-oss-120B (Apache 2.0) is OpenAI’s open-weight line. Plus Liquid AI’s LFM2.5-24B-A2B for on-device agents, and MiniMax M3 at 80.5% on SWE-bench Verified.
Composite of ten independent evaluations, read from Artificial Analysis (Index v4.3.2) on October 5, 2026. Version matters: the Index was rebased from v4.1.1, and scores across versions are not comparable.
| Rank (open weights) | Model | Index v4.3.2 | License |
|---|---|---|---|
| 1 | MiMo-V2.6-Pro — Xiaomi | 46.32 | MIT |
| 2 | GLM-5.3 — Z.ai | 44.78 | Bespoke (revenue-gated) |
| 3 | Kimi K3 — Moonshot AI | 43.59 | Bespoke |
BenchLM's open-weight composite (BenchAlign v5.8) tells the same story with different arithmetic: MiMo-V2.6-Pro first (73.7), then Qwen3.8 Max (69.7), GLM-5.3 (67.9), and DeepSeek V4.1-Flash-1017 (65.6).
The commercial frontier sits at ~95–96% on this benchmark — vendor figures on a saturated board. The open leaders, on the standard public boards:
| Model | Score | Board / note |
|---|---|---|
| DeepSeek-V4-Pro-Max | 80.6% | Top open score, SWE-bench Verified (llm-stats) |
| MiniMax M3 | 80.5% | SWE-bench Verified |
| Mistral Medium 3.5 | 77.6% | Developer-reported figure |
| Kimi K2 Instruct | 65.8% | SWE-bench Verified, Agentic Coding board (mini-SWE-agent harness) |
| GLM-5.2 | 62.1% | SWE-bench Pro — above GPT-5.5's 58.6% on the same board |
Read with care: the harness alone moves SWE-bench results by an estimated 5–15 points, and Verified is flagged contaminated/saturated by the leaderboard operators themselves. These numbers rank models inside one harness at one date — they are receipts, not verdicts.
"Open" is doing heavy lifting in 2026. The weights may be downloadable while the license still gates commercial use — the Watch records the license next to every flagship:
| License | Models | What it means |
|---|---|---|
| MIT | DeepSeek V4 line · GLM-5.3-Flash · MiMo-V2.6-Pro | Download, fine-tune, deploy commercially, no royalties. |
| Apache 2.0 | Qwen3.8-27B · Gemma 4 · gpt-oss-120B · Muse Glimmer 30B | Permissive, with patent grant; standard open-source terms. |
| Bespoke / gated | GLM-5.3 · Kimi K3 · Qwen flagships · Llama 4 (Community License) | Weights are public, but terms add conditions — revenue gates, hosting restrictions, or acceptable-use clauses. Read before you ship. |
Dated, sourced, and limited to what actually shipped.
Open weights, datasets, and training recipes. Bolt uses 9B of its 74B parameters per token and, per NVIDIA, fits 24GB of memory at native 2-bit quantization; a 4GB-class Bolt E4B is scheduled for Oct 23. Throughput and ranking claims are NVIDIA's own.
24B total parameters, ~2B active per token, 131K context, trained inside agent harnesses for multi-step tool calling. Base and post-trained weights are on Hugging Face (GGUF, MLX, ONNX, FP8 builds).
1.05T-parameter MoE, 49B active, 1M-token context, native vision. API is live ($1.36 / $4.18 per 1M tokens); the open weights ship end of October, so self-hosting is not possible yet.
How this page works: standings are snapshots with their read dates shown; vendor-reported figures are labeled as such; when a leaderboard rebases its index, the old numbers are retired rather than mixed in. Updated as the open frontier moves.