best 8 coding llms

apache 2.0, and we read the licence file rather than trusting the model-card badge — no revenue cap, no use restrictions, no naming mandate. after codestral's non-production bar and devstral's revenue cliff, that plainness is the argument.

262,144 tokens natively, extendable to a million through yarn, which is what makes repository-scale agentic coding practical rather than aspirational. the flagship is 480b total with 35b active.

the flagship also needs about 240gb at 4-bit, which is multi-gpu server territory. qwen3-coder-next, added in february 2026 at 80b total and 3b active, is the answer to that — alibaba claims it matches models with ten to twenty times the active parameters, and at roughly 40gb it lands within reach of serious local hardware. one caveat: secondary sources report swe-bench verified anywhere between 50 and 66.5% and we found no authoritative primary figure, so we're treating that metric as unresolved.

deepseek moved in the opposite direction to everyone else in this ranking. the original 2023 coder generation paired mit code with a separate, more restrictive custom weights licence; the current v4 weights are plain mit for both. in a field where minimax relicensed to non-commercial after launch and mistral gates its best coder behind a revenue cliff, a vendor loosening its terms deserves saying out loud.

the performance is the top of this page: 80.6% swe-bench verified for v4-pro and 79.0% for v4-flash, with 55.2% on livecodebench in flash's non-thinking mode. these are aggregated figures rather than ones we confirmed on deepseek's own benchmark page.

what's gone is the thing that made deepseek coder useful to individuals. the standalone line no longer exists — 'deepseek-coder' is an api alias pointing at the general v4 family — and the smallest current option, v4-flash, is 284b total and needs about 140gb at 4-bit. the original shipped at 1.3b to 33b. anyone who chose it for local editor completion has no successor here.

bytedance trained it natively on fill-in-the-middle and suffix-prefix-middle objectives, which is what editor integration actually needs and what most of the large agentic models in this ranking don't offer. at 8b and roughly 5gb quantised it runs on hardware you already own, and completion latency is local rather than a round trip.

the licence is plain mit, confirmed from the repository. no thresholds, no use restrictions, no attribution mandate — the same standing as qwen3-coder and deepseek, at a fraction of the size.

it reports humaneval 84.8, mbpp 85.2 and 24.7 on livecodebench for the october 2024 to february 2025 window, self-reported on its own model card. the ceiling is real: 8b is the only size, there's been no update since june 2025, and it cannot compete with 480b agentic models on repository-scale work. within the small-model tier it is the strongest thing here.

apache 2.0, verified in the repository's raw licence file, with ibm's documentation and the licence in complete agreement — which after codestral and starcoder2 is worth noting as a category in itself. granite 4.1 also signs its weights cryptographically, which nothing else here does.

the practical specs are strong for local work: 8b at roughly 5gb quantised, native fill-in-the-middle documented by ibm, and context extended to 512k in 4.1 through staged extension with model merging. for an editor integration inside a regulated environment that combination is hard to beat.

the caveat is what happened to the specialist. ibm's own model card for granite code says 'not recommended for use in new projects' — the dedicated line is deprecated and coding is now one capability of the general granite models. that means cleaner licensing and less code-specific tuning, so it trails dedicated coders like qwen3-coder and seed-coder on pure code benchmarks.

devstral 2's 123b model carries a modified mit licence that revokes all rights if your company's global consolidated monthly revenue exceeded $20 million in the preceding month. that is a hard cutoff rather than a graduated fee, pushing larger firms onto mistral's paid api or a negotiated deal. devstral small at 24b is plain apache 2.0 with nothing attached.

credit where it's due: mistral's own launch material states the split explicitly rather than letting people discover it in the licence file, which is the failure mode we found at fish speech and marker. same company shipped codestral under a flat non-commercial bar while framing it as a download, so the contrast within one vendor is instructive.

on capability, devstral is built for agentic coding scaffolds rather than editor completion — it was developed with all hands ai for openhands-style tooling — and devstral small 2507 reported 53.6% on swe-bench verified, up 6.8 points on its predecessor. the 24b runs in about 14gb; the 123b needs 65 to 70gb.

glm-4.6 was zhipu's coding-focused release and is reported at near-parity with claude sonnet 4 on cc-bench, a human-evaluated real-world coding benchmark. it's positioned for claude code and cline-style integrations, with 200k context, up from 128k in glm-4.5.

the licence is the problem for a page built on reading licence files. hugging face tags it mit and zhipu's messaging says fully open and self-hostable with no lock-in, but the raw licence file returned 404 on both paths we tried, so we could not check the text for an added clause. given that kimi k3's licence carried two commercial gates invisible from its badge, a metadata tag is not proof.

at roughly 355b it needs about 180gb at 4-bit — datacentre only. and the coding specialism has moved on: glm-5.1 and 5.2 have shipped as general flagships claiming stronger coding, neither branded as a coder variant, so 4.6 is the last one carrying the label.

trained on the stack v2 across more than 600 programming languages, with native fill-in-the-middle, at sizes from 3b to 15b — roughly 2gb to 9gb quantised. bigcode's paper reports the 15b matching or beating codellama-34b, a model more than twice its size. for breadth of language support at local scale nothing here competes.

the licence is openrail-m and it is not what people assume. commercial use is permitted, and attachment a attaches binding use restrictions you must pass into your own terms of service for your end users: no malware, no verifiably false or harmful disinformation, no impersonation for deception, no undisclosed fully-automated decisions affecting people's legal rights, and public-facing generated content must intelligibly disclose that it is machine-generated. the model card doesn't spell those out — the attachment does.

the age is the other issue. released february 2024 with no successor identified, it now sits well behind current-generation models on agentic and swe-bench-style tasks, and its 16k context with 4k sliding-window attention is short for repository work.

the mistral ai non-production licence bars commercial and production use outright for the open-weight release — research and testing purposes only. that is stricter than every threshold-based licence here: devstral's $20m revenue cliff, kimi k3's scale gates and marker's funding cap all permit shipping until you get large. this permits none.

the framing compounds it. mistral's launch page emphasises downloading from hugging face without leading with the prohibition, and the same company ships devstral small under plain apache 2.0 and states devstral 2's revenue cap explicitly. the candour is inconsistent within one vendor's own product line.

there's a second problem for anyone assessing it today. the current codestral 25.08 appears to be offered only through mistral's paid api, with no open-weight release we could find, so the open-weight framing may now apply only to the legacy 22b checkpoint from may 2024. the fill-in-the-middle support remains genuinely good for editor work — if you're paying for the api, which is the point.