Best Open-Source AI Coding Model.

the ranking

the top picks, decoded

1. Kimi K3 — 93.4% SWE-bench Verified

The highest-scoring downloadable coding model we track, and by far the cheapest way to buy a 90%+ result at $3 per 1M input. Weights shipped Jul 27, 2026 as 96 shards totalling 1.56 TB, 4-bit MXFP4 only: roughly 1,454 GiB, which fits one eight-card 192GB node. No full-precision base checkpoint was published. Independent (vals.ai, mini-swe-agent bash-only harness): SWE-bench Verified 93.40% ±1.11. Verified Jul 18, 2026 — vals.ai had not run K3 at our Jul 17 check and now ranks it third overall (above GPT-5.6 Luna at 93.0% and Claude Opus 4.8 at 88.6%), making it the highest-scoring open-weight coder we track. Moonshot published no SWE-bench Verified score of its own, only Terminal-Bench 2.1 88.3% on its KimiCode harness, so we rank on the independent number per our standing rule. Corroborated by a second independent evaluator: Artificial Analysis, which runs its own tests, has K3 4th on GDPval-AA v2 (Elo 1684), 2nd on AA-Briefcase (Elo 1545) and 4th of 187 on its Intelligence Index v4.1 (57.1). It also classifies K3 as proprietary outright. AA measured Terminal-Bench 2.1 at 85.02% against Moonshot's own claim of 88.3%, a 3.3-point vendor overstatement, which is why we print the independent figure. Blind developer voting on Arena WebDev ranks it 1st of 99. Price confirmed on Moonshot's own pricing page: $3/$15 per 1M, cache-hit input $0.30 (platform.kimi.ai, checked Jul 19, 2026). **open: false, corrected Jul 19, 2026** — we had this flagged as open weights, which put it top of our "best open-source coding model" page as a model nobody can download. Checked Jul 19: no K3 repo under huggingface.co/moonshotai (18 repos, none K3) and github.com/MoonshotAI/Kimi-K3 returns 404. Moonshot's own blog still says "The full model weights will be released by July 27, 2026." Until they actually ship, open is a roadmap item, not a fact, and this row is an API model. Flip back to open: true when the weights land. Re-checked Jul 27, 2026 (Moonshot's own deadline day): huggingface.co/moonshotai still lists no K3 repo, newest is Kimi-K2.7-Code. Score unchanged at 93.40% on vals.ai. **open: true, restored Jul 27, 2026 13:31 UTC**: the weights actually landed at huggingface.co/moonshotai/Kimi-K3 later the same day, 96 safetensors shards, 1,560,998,983,759 bytes (about 1,454 GiB), under a bespoke "Kimi K3 License" that is permissive but not Apache 2.0 (separate agreement required above $20M MaaS revenue; attribution required above 100M MAU). Verified against the Hugging Face API and config.json, not a press report. One caveat that belongs on the record: the release is 4-bit MXFP4 only for the routed experts (quantization-aware trained from the SFT stage, format mxfp4-pack-quantized, group size 32), with attention, shared experts, LM head and vision tower left at bf16, and there is no base checkpoint, so this is open for inference rather than open for training. Score unchanged at 93.40%; nothing was re-ranked by this flag. See /p/kimi-k3-weights-landed-4-bit-only/.

2. DeepSeek V4 Flash — 88.8% SWE-bench Verified

The cheapest model making a near-frontier agentic claim: 284B MoE with 13B active, 1M context, and native Responses API plus Codex support at $0.28 per 1M output. Independent (vals.ai, Aug 5 2026, mini-swe-agent bash-only harness): SWE-bench Verified 88.8%, for the DeepSeek-V4-Flash-0731 build, matching the checkpoint DeepSeek own change log describes. Ranks above Claude Opus 4.8 (88.6%) and below GPT-5.6 Luna (93.0%). Vendor-reported agent numbers for the same build (DeepSeek change log, Jul 31 2026): Terminal-Bench 2.1 82.7, Cybergym 76.7, Toolathlon verified 70.3, DeepSWE 54.4, NL2Repo 54.2, Agent Last Exam 25.2, produced with DeepSeek own unreleased "DeepSeek Harness minimal mode" and not independently reproduced. The V4-Flash-Preview model card (pre-0731 weights) separately listed SWE-bench Verified 79.0 and Terminal-Bench 2.0 56.9; those are not carried over. Same architecture and size as the preview (284B total / 13B active, FP4+FP8), re-post-trained only. Not to be confused with DeepSeek V4 Pro (ranked, 80.6%) or the plain DeepSeek V4 (77.4%). See /p/deepseek-v4-flash-claims-82-7-terminal-bench/.

3. GLM 5.2 — 82.8% SWE-bench Verified

The best open-weight coder on this board that you can actually download today: an independently measured 82.8%, MIT-licensed, and roughly a third the input price of the closed models above it. Independent (vals.ai, Jul 22 2026, mini-swe-agent bash-only harness): SWE-bench Verified 82.8% ±1.69, 9th of 75 systems. Added Jul 27, 2026 after a routine sweep found vals.ai had scored it and we were not carrying it. Z.ai published no SWE-bench Verified figure of its own, so the independent number is the ranked score with no vendor claim to disclose against it. This is now the highest-scoring genuinely downloadable model on the board: Kimi K3 scores higher at 93.4% but its weights remain unpublished, so K3 is listed as closed. Weights went up on Hugging Face and ModelScope under an unrestricted MIT license on Jun 16, 2026. Priced $1.40/$4.40 per 1M, cached input $0.26. vals.ai measured $0.71 per SWE-bench task, cheaper per solved task than every closed model ranked above it except GPT-5.6 Luna. Context 1M, max output 131k.

how we rank

We rank by SWE-bench Verified (500 real, human-validated GitHub issues resolved end-to-end), tiebroken by the harder SWE-bench Pro. A score is only printed once confirmed against an independent evaluation or the maker's primary source — and every row states which kind it is. Where both exist, we print both: one as the ranked score, the other in that row's note. We would rather show you the gap than ask you to trust our pick. Our independent reference is vals.ai, which runs every model itself through the same minimal bash-only harness (mini-swe-agent), so the models are compared on equal footing. That matters more than it sounds: SWE-bench scores a model and its scaffolding together, and vendors report using their own tuned scaffolds. Against vals.ai's neutral harness, the vendor claims on this board run 2.6 to 11.6 points optimistic. So rows marked vendor-reported are best-case numbers and are not strictly comparable to the independent ones — where we know the independent figure, we print it in the row's note. That is a deliberate choice and you should know we made it: on four rows (Claude Sonnet 5, MiniMax M3, Qwen3.7 Max, Kimi K2.6) an independent score for that exact model exists and is lower, and we still rank on the maker's published figure because it is the number that model is sold and quoted on. We disclose the independent one beside it instead of quietly restating the board on a single evaluator's harness choice. The honest consequence: positions that straddle the two regimes are approximate. Qwen3.7 Max is the sharpest case — it sits at #16 on Alibaba's 80.4%, and on the neutral harness its 68.8% would put it far down the table. Note that llm-stats, which we previously miscredited as an independent tracker, labels its own SWE-bench Verified table "Verified: 0 / Self-reported: 104"; it aggregates vendor claims. Models still being checked are marked “verifying” and shown without a number rather than estimated. Prices are per 1M input tokens on the standard API tier and can change — always confirm current pricing with the provider.

Want the raw numbers? The full dataset is public: JSON · CSV.

From our full AI Coding Leaderboard (2026-08-13). We only rank scores confirmed against primary sources.