Top Open Source LLMs to Watch in 2026

Current State of Open Source LLMs

Open weights now cover the whole range, from a 30B on one workstation card up to a 2.8T mixture of experts. How far behind the frontier they sit depends on which board you open and which version of it: four points on Mozilla's July reading, twelve on the current Artificial Analysis index, fourteen on SWE-bench Verified.

Price moves fastest, and since August it has started to split by direction. Input stayed cheap. On 7 October 2026 gpt-oss-120b listed at $0.037 per million input tokens and DeepSeek V4 Flash 0731 at $0.018, against $4.00 for Claude Opus 5.5, which Anthropic released on 22 September at $4 in and $20 out. Output is where it gets complicated. OpenRouter's headline for DeepSeek V4 Flash now pairs $0.03 in with $1.28 out, about 7.6 times the $0.168 it listed in August, yet one provider serves the same weights at $0.084 out.

At the top of the open range, Kimi K3 lists at $0.62 in and $15.00 out: about a sixth of Opus 5.5 on input and three quarters of it on output. Whether that saves money depends almost entirely on how much your workload writes compared with how much it reads. That ratio, more than any single price, is the number to carry into a build-versus-buy sum.

The Artificial Analysis index averages a broad basket of evaluations; SWE-bench Verified measures one hard thing, resolving real GitHub issues, and agentic work lives in exactly the tail that averaging hides.

Carry the version string too. Between v4.1.1 in August and v4.3.2 in October, Artificial Analysis rescored every model: Kimi K3 went from 60 to 44 and Qwen3.8-27B from 52 to 34. Over the same weeks Claude Opus 5.5 arrived at 58, the highest score the index has measured. So a 3-point gap became a 12-point one, and none of that means an open model got worse. A screenshot of a leaderboard without its version is close to useless a month later.

DeepSeek-V4-Pro-Max, the top open-weight entry on SWE-bench Verified at 0.806 on 7 October, has no visible repository under DeepSeek's Hugging Face organisation: a board score, not a download. MiniMax M3, which you can download, sits right behind it at 0.805.

Licenses are the third axis, and they stopped travelling with the family name. The cleanest example arrived in late August. GLM-5.3-Flash went out on 26 August under plain MIT, and two days later GLM-5.3 itself, 753B parameters, under a custom glm-5.3 license. It reads like MIT, except that model-as-a-service operators with more than $10 billion in revenue over any twelve months must first pass a Z.ai security review, and the license publishes no criteria or timeline for that review.

Qwen does the same: the 2.4T flagship carries a custom qwen3.8-max string while Qwen3.8-27B, two days later, is plain Apache 2.0. Gemma 4 moved to Apache 2.0, replacing the old Gemma Terms, while the Llama 4 Community License still requires a separate commercial license above 700 million monthly active users and incorporates an acceptable use policy by reference. Read the file, not the family page.

Meta shipped Muse Glimmer on 10 August under Apache 2.0 and the MobileMoE on-device family on 26 August under a noncommercial research license. Same vendor, same month, opposite terms.

A family name no longer tells you whether there is anything to download, either. Mistral Large 4 has been on OpenRouter since 6 October, but Mistral's Hugging Face organisation still tops out at Large 3, so as of 7 October we could not confirm open weights for it. Qwen3.8 Max and Max Prime, the latter listed on 23 September, are API-only, and so is Meta's Muse Spark 1.3. Check the Hugging Face organisation before you plan around a self-hosted version.

Key Trends in Open Source LLM News

Two trends carry most of the open source llm news since December 2025: models grew at the top and shrank at the bottom at once, and the serving stack stopped lagging.

Latest Model Releases. This is the llm model release news that changed something between December 2025 and today. The date is the weights date wherever it differs from the announcement, because a model you cannot download is not a release.

Parameter count stopped predicting rank. Qwen3.8-27B, a dense 27B, scores 34 on Artificial Analysis v4.3.2, ahead of the 428B MiniMax-M3 at 29 and the 975B Inkling at 25. The Llama line is still missing, with no new Llama weights since April 2025; Meta's open release this year came under the Muse name instead.

Innovations in Tooling. Engine support now arrives with the model instead of a month later. Kimi K3's card names vLLM, SGLang and TokenSpeed, so day-zero support is arranged before shipping. Quantization moved with it: labs publish MXFP4 weights natively, and llama.cpp gained NVFP4 kernels for Blackwell in April.

Noteworthy Open Source LLMs

Five models cover most of what a team would shortlist, each with its exact license string. Prices are OpenRouter's, retrieved 7 October 2026.

1. GLM-5.3. Z.ai announced it in mid-August and published the weights on 28 August 2026: 753B parameters, with 88.2 on Terminal-Bench 2.1 per its card and 45 on Artificial Analysis v4.3.2, second among open models. The license is the custom glm-5.3 described above, MIT in all but the security-review clause for the very largest hosting companies. Its smaller sibling, GLM-5.3-Flash at 320B, came out two days earlier under plain MIT and scores 42. The price shows the same split as DeepSeek's: a headline of $0.07 in and $7.00 out, and a route at $0.42 in and $1.32 out.

2. MiniMax-M3. Released 1 June 2026: 428B total, 23B active, a 1M context window, and 80.5 on SWE-bench Verified per its card. The license string is `minimax-community`, neither Apache nor MIT. At a headline $0.30 in with $1.20 out, or $0.23 and $0.96 on the cheapest route with a 262K context, it is still one of the cheapest capable options on the board.

3. Kimi-K3. Moonshot's flagship, weights on 26 July 2026: 2.8T total, 104B active, 896 experts, MXFP4 weights trained with quantization-aware training. Its card reports GPQA-Diamond 93.5, and it scores 44 on Artificial Analysis v4.3.2, third among open models after leading the column at 60 on the August version. The license is a custom "Kimi K3 License" covering code and weights together. At $0.62 in with $15.00 out it is a pricing shape rather than a discount: an agent that reads a lot and writes little saves real money against Opus 5.5, while long-form generation pays close to frontier rates for a model 14 points behind it on that index.

4. DeepSeek-V4. The line splits in two, and the split is the story. V4 Pro 0813 is MIT, 1.6T total with 49B active, and scores 87.9 on Terminal Bench 2.1, but its checkpoint runs to roughly 893 GB, so MIT gets you a model an 8xH100 node cannot hold. V4 Flash 0731, same license, is the realistic one: 284B total, 13B active, 82.7 on Terminal Bench 2.1. Its headline is now $0.018 in and $1.28 out, while the cheapest output route charges $0.044 and $0.132.

The newer option is V4.1 Flash, released on 10 September under MIT. Its card describes a 552B backbone that activates 8B parameters per token in prefill and 16B in decode, native image input and a 1M context. It picked up about 1.26 million downloads in its first 30 days, scores 39 on Artificial Analysis, fifth among open models, and lists at $0.05 in and $1.20 out. A separate V4 Flash Vision Exp checkpoint appeared on 31 August, also under MIT.

5. MiMo-V2.6. Xiaomi published the family on 21 September 2026 under MIT. MiMo-V2.6-Pro is 1.02T total with 42B active, takes text, images, video and audio natively, and has a 1M context window; a Flash model and a 9B distilled variant sit below it. It is the top open model on Artificial Analysis v4.3.2 at 46. On OpenRouter it lists at $0.435 in and $0.87 out, but through only four providers on 7 October, so there are fewer routes to shop between than for the DeepSeek line.

Open Source LLM Tooling News Today

If you want open source llm tools news today rather than a curated tool list, the release notes are the feed, with versions, dates and numbers attached.

Ollama shipped Muse Glimmer on day zero on 10 August with initial MLX support on Apple Silicon, then in v0.32.15 on 19 August cut time to first token roughly in half by caching model metadata. Since then, v0.40.0 on 25 September made MLX the default runtime on Apple Silicon. The 0.35 releases on 28 and 29 September added "decision models" behind a new /v1/systemone endpoint, which return choices and probabilities instead of text - a better fit for routing and ticket triage than prompting a chat model to answer with one word.

LM Studio reads the same way: MTP speculative decoding stable in 0.4.14 on 22 May, then internal-network model endpoints and local server API keys in 0.4.21 on 12 August. Version 0.4.22 on 28 August added support for DFlash, DSpark and MTP drafters, 0.4.24 on 9 September brought advanced llama.cpp argument overrides for GGUF loading, and 0.4.25 on 19 September added the Splash engine on Apple Silicon M3 or newer.

llama.cpp picked up native NVFP4 for Blackwell in a PR merged on 28 April 2026, and GGML_TYPE_NVFP4 = 40 now sits in the ggml type table. Published throughput figures for it trace back to forum arithmetic, so measure it yourself.

Most open source llm tooling news on the server side is feature work rather than headlines, like the prefill/decode-aware router vLLM added in December 2025. For numbers that go stale in weeks, SemiAnalysis runs InferenceX with weekly snapshots.

Advancements in Inference Techniques

Enhancements in Performance. Most open source llm inference news is throughput. vLLM reported 2,200 output tokens per second per H200 serving DeepSeek on 17 December 2025, and on GB200 NVL72 in February 2026 measured 26.2K prefill and 10.1K decode tokens per GPU per second, three to five times H200.

SGLang's chunked pipeline parallelism, 15 January 2026, is the long-context one: prefill throughput up 3.31x on six nodes of eight H20s and time to first token down 67.9% on DeepSeek-V3.1. In the same post a one-million-token prompt reached its first token in 420.91 seconds, which makes a 1M window a batch feature rather than a chat feature.

Speculative decoding now ships from the labs: DeepSeek's V4 checkpoints carry a DSpark module, and Unsloth measured V4 Flash at 120 tokens per second with it on a B200 against 60 without.

Compatibility with Existing Systems. Formats are where this meets your existing hardware, and they are not free. A benchmark of post-training quantization under microscaling formats, January 2026, found MXFP8 near-lossless while MXFP4 degraded accuracy substantially, which matters because MXFP4 is what several labs now ship natively.

Unsloth publishes measured footprints and divergence numbers for its DeepSeek V4 GGUF builds, and two things follow. The 4-bit build is barely smaller than the 8-bit one, because the experts already ship in MXFP4 and only non-expert tensors get requantized. And 4-bit is not free: top-token agreement drops to 96.28%, so roughly one token in 27 differs from full precision. Invisible in one answer, very visible in a forty-step agent loop, which is a separate problem worth measuring.

Then the arithmetic that decides whether it applies to you. V4 Pro's 893 GB against 640 GB in an 8xH100 node means the weights do not fit before a single token of KV cache. V4 Flash lands nearer 170 GB at native FP4 and FP8, two H200s, or 90 to 100 GB at community INT4. gpt-oss-120b fits one 80 GB H100. Running any of these on your own hardware turns on total memory rather than VRAM alone, which is the sizing question worked through per quant build.

Community Contributions and Collaborations

Download counts show where the community went. Hugging Face's 2026 open model report, as reported by China Daily on 16 August 2026, puts Qwen at 2.045 billion downloads in 2026 against 418 million for Google and 227 million for Meta, with 151,448 models derived from Qwen. Gemma has passed 400 million downloads and 100,000 community variants.

Ai2's OLMo 3 is a different kind of contribution: it shipped in November 2025 with the training data, the code and every intermediate checkpoint next to the weights, which is what lets somebody reproduce a result rather than re-run it.

Collaboration also shows up as a launch checklist item: naming vLLM, SGLang or llama.cpp on a model card is a commitment made before release, and its absence is a real cost for whoever adopts first.

The newest collaboration is a license rather than a model. NVIDIA's Nemotron 3.5 Lightning and Poolside's Laguna S 2.1, a 118B mixture of experts released on 21 July 2026, both ship under `openmdw-1.1`, the Linux Foundation's model-distribution license picking up its second and third adopters while the OSI question stays open.

Future Outlook for Open Source LLMs

Epoch AI converts the gap into time: since January 2026 the strongest open-weight models have lagged frontier closed models by an average of four months, or eight ECI points. Epoch makes no claim about that lag widening or narrowing, and neither will we. Four months is short enough that waiting for the open version is a real option.

The event we were watching in August has resolved. GLM-5.3's weights landed on 28 August, and the answer to "on what terms" was neither MIT nor an API-only release but something in between: MIT with a carve-out aimed at the handful of hosting companies big enough to compete with Z.ai's own API. It costs a mid-sized company nothing today. It also means the license-per-checkpoint habit is no longer optional, because two models released two days apart under one family name now carry different terms, and we would not be surprised to see that shape again from other labs.

Safety review has become a shipping-schedule variable. TechCrunch reported on 4 August 2026 on a SaferAI evaluation in which GLM-5.2, tested through Z.ai's public API, refused none of the offensive cyber or biology tasks, while Claude Opus 4.7 refused so consistently that the evaluation could not be completed. It cuts both ways: Far.ai keeps a leaderboard of the universal jailbreaks it found in closed frontier models.

The hardware shipped, mostly. Apple announced M5 Max and M5 Ultra Mac Studio configurations on 25 August 2026, up to 512 GB of unified memory and 1.2 TB/s of bandwidth, and the machines went on sale on 22 September. The 512 GB configuration is the one that matters for large open models, and Macworld reports it will not be available until October; we have not seen a confirmed ship date. Apple names no parameter count for the frontier-class open-weight models it says the machine runs.

The genuinely open question is who ends up using these models. Menlo Ventures, from November 2025 fieldwork with 495 US enterprise AI decision-makers, puts open-source models at 11% of enterprise LLM API usage, down from 19% a year earlier, with no 2026 update since. Mozilla's State of Open Source AI from July 2026 has open weights going from about a third of OpenRouter tokens in late 2025 to a majority by mid-2026.

The two count different people: Menlo counts enterprise procurement, Mozilla routed developer traffic. The bridge is inside Mozilla's own numbers: 79% of developers adding AI use open models, but only 53% of open adopters reach production against 63% for closed.

Five checks to run before you commit to an open model

About a week of work, none of it needing a vendor.

Open the LICENSE file on the exact checkpoint. Look for a monthly-active-user threshold, an attribution string you must display, a naming rule on derived models, and any policy incorporated by reference.

Price a real week of traffic, and write the date next to the total. Input and output separately, at today's rates, on the provider route you will actually use. On 7 October the headline and the cheapest output route for DeepSeek V4 Flash differed about fifteenfold on output, and prices here have a shelf life of weeks.

Size the memory before you read a benchmark. Parameters times bytes per parameter, plus KV cache for your context length and concurrency. If the answer is a cluster, the benchmark was never the question.

Evaluate the quantization you will deploy, on your own questions. One to two hundred real queries with known-good answers, against the exact build at the exact context length. Published divergence numbers hold in aggregate, not on your prompts.

Check that the model card names your serving engine. Day-zero support in vLLM, SGLang or llama.cpp is the difference between a two-day deployment and a two-month one. Then bookmark a live source: llm-stats updates daily, InferenceX weekly.

One concrete thing this week: take your top three candidates, open the LICENSE file on each, and price a real week of traffic with the date written down. The model is rarely the bottleneck anyway, and getting one to your own data safely usually costs more engineering than the choice between checkpoints. If you would rather someone ran that against your workload, that is what the free audit is for.

FAQ