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, twenty-six on the current Artificial Analysis index if you count only weights you can download, about thirteen on SWE-bench Verified.
Price moves fastest, and it splits by direction. Input stayed cheap. On 29 October 2026 the cheapest input route in our table was gpt-oss-120b at $0.03 per million tokens, and DeepSeek's newest Flash, V4.1 Flash 1017, listed at $0.15 in and $0.60 out, against $4.00 and $20.00 for Claude Opus 5.5, which Anthropic released on 22 September. DeepSeek's own V4 Flash endpoint is gone: DeepSeek's API now answers every Flash request, including the retired deepseek-v4-flash name, with V4.1-Flash-1017 at $0.15 / $0.60 off-peak and $0.30 / $1.20 at peak. Output is where it gets complicated, and the route matters more than the model.
At the top of the open range, Kimi K3's headline is $0.239 in and $10.80 out: about 6% of Opus 5.5 on input and 54% of it on output. Its successor, Kimi K3.1 Pro, is API-only for now at $2.00 in and $10.00 out from Moonshot. Whether either 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.
| Model | Weights | Headline in / out | Cheapest-output route | Context | Providers |
|---|---|---|---|---|---|
| gpt-oss-120b | Open | 0.15 / 0.60 | 0.03 / 0.15 | 131K | 25 |
| DeepSeek V4 Flash | Open | 0.091 / 0.182 | 0.05 / 0.10 | 1M | 17 |
| DeepSeek V4 Flash 0731 | Open | 0.119 / 0.238 | 0.057 / 0.134 | 1M | 25 |
| DeepSeek V4.1 Flash | Open | 0.045 / 0.50 | 0.045 / 0.179 | 1M | 41 |
| DeepSeek V4.1 Flash 1017 | Open | 0.15 / 0.60 | 0.101 / 0.15 | 1M | 42 |
| MiniMax M3 | Open | 0.30 / 1.20 | 0.224 / 0.88 (524K context) | 1M | 13 |
| MiniMax M3.1 Flash | Open | 0.15 / 0.60 | 0.135 / 0.54 | 1M | 13 |
| MiMo-V2.6-Pro | Open | 0.435 / 0.87 | 0.37 / 0.74 | 1.05M | 5 |
| Qwen3.8-27B | Open | 0.039 / 1.25 | 0.039 / 1.25 (262K context) | 1M | 21 |
| Beam 1.0 | Open | 0.25 / 1.00 | 0.25 / 1.00 | 1M | 5 |
| DeepSeek V4 Pro 0813 | Open | 0.66 / 1.98 | 0.45 / 0.80 | 1M | 26 |
| GLM-5.3 | Open | 0.07 / 1.40 | 0.28 / 0.88 | 1M | 49 |
| Muse Spark Open | Open | 1.25 / 4.25 | 1.25 / 4.25 | 1M | 6 |
| Kimi K3 | Open | 0.239 / 10.80 | 1.65 / 8.25 | 1M | 30 |
| Kimi K3.1 Pro | Due by 4 Nov | 2.00 / 10.00 | 2.00 / 10.00 | 1M | 2 |
| Claude Opus 5.5 | Closed | 4.00 / 20.00 | 4.00 / 20.00 | 1M | 11 |
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. Artificial Analysis has now rescored everything twice since August. Between v4.1.1 and v4.3.2 Kimi K3 went from 60 to 44; between v4.3.2 on 7 October and v5.0 on 29 October, MiMo-V2.6-Pro went from 46 to 30, GLM-5.3 from 45 to 32 and Qwen3.8-27B from 34 to 20, while Claude Opus 5.5 stayed at 58 (Google's Gemini 4 Carbon now edges it at 59 on the same board). The 12-point gap we printed on 7 October is 26 today for weights you can download, or 12 if you count Kimi K3.1 Pro at 46, whose weights Moonshot has promised but not published. None of that means an open model got worse. A screenshot of a leaderboard without its version is close to useless a month later.
On SWE-bench Verified the top open-weight entry on 29 October is Reflection's Beam 1.0 at 0.816, which you can download under Apache 2.0. Behind it sit DeepSeek-V4-Pro-Max at 0.806, still with no visible repository under DeepSeek's Hugging Face organisation, and MiniMax M3 at 0.805. The closed reference on that board is still Claude Fable 5 at 0.950; llm-stats has no entry yet for Claude Fable 5.1.
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. Then on 15 October came Muse Spark Open, 816B total with 42B active, under a third set of terms: the custom Muse Model License, which requires a visible "Built with Muse Spark Open" notice and, per the model card, a separate written agreement with Meta before you offer it as a model-as-a-service once revenue passes $20 million in any twelve months. Access on Hugging Face is gated and approved manually.
A family name no longer tells you whether there is anything to download, either. Mistral Large 4 has been on OpenRouter since 6 October, and Mistral's launch post promises the weights "by the end of the month", but on 29 October Mistral's Hugging Face organisation still tops out at Large 3 and no license has been named. Kimi K3.1 Pro, launched on 23 October, is API-only today; Moonshot says the full weights "will be released by Nov 4, 2026" and has not named a license. Qwen3.8 Max and Max Prime, the latter listed on 23 September, are API-only, and so is Meta's hosted Muse Spark 1.3; Muse Spark Open is a separate checkpoint. 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.
| Model | Released | Total / active | License | Benchmark |
|---|---|---|---|---|
| Mistral Large 3 | 2 Dec 2025 | 675B / 41B MoE | Apache 2.0 | - |
| Gemma 4 | 2 Apr 2026 (announced) | 31B dense, 26B-A4B, 12B, E2B/E4B | Apache 2.0 | MMLU-Pro 85.2 (31B) |
| MiMo-V2.5-Pro | 22 Apr 2026 | 1.023T / 42B | MIT | - |
| DeepSeek V4 Pro Preview + V4 Flash | 24 Apr 2026 | 1.6T / 49B; 284B / 13B | MIT | Terminal Bench 2.1 72.1 |
| IBM Granite 4.1 | 29 Apr 2026 | 3B / 8B / 30B dense | Apache 2.0 | 5.33% WER (Speech 2B) |
| Cohere Command A+ | 20 May 2026 | 218B / 25B | Apache 2.0 | - |
| MiniMax M3 | 1 Jun 2026 | 428B / 23B | minimax-community | AA Index 19 |
| NVIDIA Nemotron 3 Ultra | 4 Jun 2026 | 550B MoE | Nemotron Open Model License | AA Index 12 |
| GLM-5.2 | weights 16 Jun 2026 | ~744B / ~40B | MIT | - |
| Inkling / Inkling-Small | 15 Jul 2026 | 975B / 41B; 276B / 12B | Apache 2.0 | AA Index 12 |
| Poolside Laguna S 2.1 | 21 Jul 2026 | 118B / ~8B MoE | openmdw-1.1 | Terminal-Bench 2.1 70.2 |
| Kimi K3 | weights 26 Jul 2026 | 2.8T / 104B, 896 experts | Kimi K3 License (custom) | GPQA-Diamond 93.5 |
| DeepSeek V4 Flash 0731 | 31 Jul 2026 | 284B / 13B | MIT | Terminal Bench 2.1 82.7 |
| Muse Glimmer (Meta) | 10 Aug 2026 | 30B distilled | Apache 2.0 | AA Index 6 |
| Nemotron 3.5 Lightning | 11 Aug 2026 | 30B / 3B | openmdw-1.1 | - |
| Qwen3.8-2.4T-A95B | 12 Aug 2026 | 2.4T / 95B | qwen3.8-max (custom) | - |
| DeepSeek V4 Pro 0813 | 13 Aug 2026 | 1.6T / 49B | MIT | Terminal Bench 2.1 87.9 |
| Qwen3.8-27B | 14 Aug 2026 | 27B dense, vision | Apache 2.0 | AA Index 20 |
| GLM-5.3-Flash | 26 Aug 2026 | 320B | MIT | - |
| GLM-5.3 | weights 28 Aug 2026 | 753B | glm-5.3 (custom) | AA Index 32 |
| DeepSeek V4 Flash Vision Exp | 31 Aug 2026 | V4 Flash with vision | MIT | - |
| DeepSeek V4.1 Flash | 10 Sep 2026 | 552B backbone / 8B prefill, 16B decode | MIT | - |
| MiMo-V2.6-Pro / Flash | 21 Sep 2026 | 1.02T / 42B (Pro) | MIT | AA Index 30 |
| Muse Spark Open (Meta) | 15 Oct 2026 | 816B / 42B | Muse Model License (custom) | AA Index 25 |
| Nemotron 3.5 Bolt | 15 Oct 2026 | 74B / 9B, 2-bit QAT | openmdw-1.1 | Terminal-Bench 2.1 50.98 |
| Reflection Beam 1.0 | 16 Oct 2026 | 501B / 23B | Apache 2.0 | AA Index 18 |
| DeepSeek V4.1 Flash 1017 | 17 Oct 2026 | 552B backbone + 196B Engram | MIT | AA Index 32 |
| MiniMax M3.1 Flash | weights 23 Oct 2026 | 428B / 23B | minimax-community | AA Index 29 |
| gpt-oss-120b / 20b | 5 Aug 2025 | 117B / 5.1B, MXFP4 | Apache 2.0 | - |
Parameter count stopped predicting rank. Qwen3.8-27B, a dense 27B, scores 20 on Artificial Analysis v5.0, ahead of the 501B Beam 1.0 at 18, the 428B MiniMax-M3 at 19 and the 975B Inkling at 12. The Llama line is still missing, with no new Llama weights since April 2025; Meta's open releases this year came under the Muse name instead, Muse Glimmer in August and Muse Spark Open in October.
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 29 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 32 on Artificial Analysis v5.0, tied for first among open models you can download. 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; Artificial Analysis lists it without a v5.0 score. The price shows the usual split: a headline of $0.07 in and $1.40 out, and a route at $0.28 in and $0.88 out.
2. MiniMax-M3 and M3.1 Flash. M3 was released on 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. Its successor, MiniMax-M3.1-Flash, published weights on 23 October under the same string and the same 428B / 23B shape. It scores 29 on Artificial Analysis v5.0 against 19 for M3, and lists at $0.15 in and $0.60 out, half of M3's $0.30 and $1.20. MiniMax's card reports DeepSWE v1.1 rather than SWE-bench Verified, 73.7%, so the two are not directly comparable on coding.
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. Artificial Analysis v5.0 shows no score for K3; its successor K3.1 Pro scores 46, the highest of any open-weight family, but on 29 October it is API-only, with weights promised by 4 November and no license named yet. K3's license is a custom "Kimi K3 License" covering code and weights together. At a headline $0.239 in with $10.80 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 more than half of Opus 5.5's output rate.
4. DeepSeek-V4 and V4.1. 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 $0.119 in and $0.238 out, while the cheapest output route charges $0.057 and $0.134.
The newer option is the V4.1 Flash line. V4.1 Flash, released on 10 September under MIT, has a 552B backbone that activates 8B parameters per token in prefill and 16B in decode, plus 196B of Engram conditional memory, native image input and a 1M context. On 17 October DeepSeek followed with V4.1-Flash-1017, a minor update under the same MIT license, which DeepSeek's own API now serves for every Flash request. It ties GLM-5.3 at 32 for the top of the open board on Artificial Analysis v5.0 and lists at $0.15 in and $0.60 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 led the open board on 7 October at 46 on v4.3.2; on v5.0 it scores 30, third among open models with weights. On OpenRouter it lists at $0.435 in and $0.87 out, through only five providers on 29 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.
| Build | Size | Fidelity against full precision |
|---|---|---|
| UD-Q8_K_XL | 162 GB | Bit-identical, KL divergence ~0, 100% top-token agreement |
| UD-Q4_K_XL | 155.1 GB | KL divergence 0.0102, 96.28% top-token agreement |
| UD-IQ3_XXS | 103 GB | Sized for machines with 110-128 GB of RAM |
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. The newer V4.1-Flash-1017 is heavier: Unsloth's GGUFs run from 233 GB at 1-bit to 508.4 GB at 8-bit, and its 128 GB RAM floor holds only with the Engram table left on SSD. 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 it became orderable on 28 October, on the late-October schedule Apple had given, on the 36-core CPU, 80-core GPU M5 Ultra only. We have not seen a delivery date we could confirm. 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. Why open weights matter even to companies that never host one is covered in the open weight models explainer.
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. If those license terms are new to you, the plain-words license cheat sheet explains what each common one allows.
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 29 October GLM-5.3's headline was $0.07 in and $1.40 out, while its cheapest output route was $0.28 in and $0.88 out: cheaper on one side, dearer on the other. 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
Which open source LLM is best right now?
It depends on the benchmark, and the honest answer carries a date. On the Artificial Analysis Intelligence Index v5.0, retrieved 29 October 2026, the top open models with downloadable weights are DeepSeek V4.1 Flash 1017 and GLM-5.3 at 32, then MiMo-V2.6-Pro at 30, against 58 for Claude Opus 5.5. Kimi K3.1 Pro scores 46 but its weights are only promised for 4 November. On SWE-bench Verified the top open entry is Reflection's Beam 1.0 at 0.816. If you have to host it yourself, DeepSeek V4 Flash 0731 and Qwen3.8-27B are still the practical answers.
Is Qwen Apache 2.0?
Partly. Qwen3.8-27B, released 14 August 2026, is Apache 2.0, while the 2.4T flagship released two days earlier carries a custom qwen3.8-max license string. The license belongs to the checkpoint, not to the family name.
Can I run DeepSeek V4 locally?
Not the Pro version, unless "locally" means a cluster: the checkpoint is around 893 GB against 640 GB in an 8xH100 node, before any KV cache. V4 Flash 0731 is realistic, and Unsloth's measured GGUF builds run from 103 GB at 3-bit to 162 GB at 8-bit. The newer V4.1 Flash 1017 runs from 233 GB to 508.4 GB, with a 128 GB RAM floor only if the Engram table stays on SSD.
Are open models cheaper than the closed APIs?
On input, dramatically: on 29 October 2026 the cheapest gpt-oss-120b route charged $0.03 per million input tokens and DeepSeek V4.1 Flash 1017 listed at $0.15, against $4.00 for Claude Opus 5.5. On output it depends on the provider route: Kimi K3's headline output is $10.80, while its cheapest output route charges $8.25, against $20.00 for Opus 5.5.
- 30 October 2026Added a link to the open weight models explainer.
- 29 October 2026Benchmarks moved to Artificial Analysis Intelligence Index v5.0, which rescored every model; lead, takeaways, bars, release table and FAQ rewritten to match. Added DeepSeek V4.1 Flash 1017, Meta Muse Spark Open, Reflection Beam 1.0, NVIDIA Nemotron 3.5 Bolt and MiniMax M3.1 Flash with their license strings; noted Kimi K3.1 Pro (API-only, weights promised by 4 November) and Mistral Large 4 (weights promised for end of October, not on Hugging Face as of 29 October). OpenRouter prices restamped to 29 October; DeepSeek's first-party V4 Flash endpoint retired. SWE-bench Verified reading refreshed (Beam 1.0 now the top open entry). 512 GB M5 Ultra status updated.
- 7 October 2026Prices restamped to 7 October 2026, now with headline and cheapest-output routes. Benchmarks moved to Artificial Analysis v4.3.2, which rescored every model. Claude Opus 5.5 replaces Opus 5 as the closed reference. GLM-5.3 weights and license, MiMo-V2.6 and DeepSeek V4.1 Flash added; Ollama and LM Studio updates added; Llama 4 license link updated. SWE-bench Verified reading refreshed to 7 October.
- 31 August 2026Added a link to the DeepSeek local requirements guide from the inference section.
- 26 August 2026Prices updated to 26 August 2026. Added Poolside Laguna S 2.1 and the verified license for Nemotron 3.5 Lightning.
- 25 August 2026Published.