# MiniMax-M2.7 by MiniMax: Open-Weight Model Source: https://savrn.com/models/minimax-m2-7 Markdown alternate of the page above; the site index is https://savrn.com/llms.txt --- ## Runs On What it takes to serve MiniMax-M2.7 (228.7B parameters): the memory its weights need at each precision, and the cheapest way to rent enough data-center GPUs to hold them. | Precision | Weights | Memory needed | Cheapest setup | Per hour | Also fits | | --- | --- | --- | --- | --- | --- | | 16-bit | 457.4 GB | 548.9 GB | 2x [MI355X](https://savrn.com/ai-index/pricing/gpus/mi355x) (288 GB) Vultr | $5.18 | [3x MI300X](https://savrn.com/ai-index/pricing/gpus/mi300x) $5.55 · [3x MI325X](https://savrn.com/ai-index/pricing/gpus/mi325x) $6.00 | | 8-bit | 228.7 GB | 274.4 GB | 1x [MI355X](https://savrn.com/ai-index/pricing/gpus/mi355x) (288 GB) Vultr | $2.59 | [2x MI300X](https://savrn.com/ai-index/pricing/gpus/mi300x) $3.70 · [2x MI325X](https://savrn.com/ai-index/pricing/gpus/mi325x) $4.00 | | 4-bit | 114.3 GB | 137.2 GB | 1x [MI300X](https://savrn.com/ai-index/pricing/gpus/mi300x) (192 GB) Vultr | $1.85 | [1x MI325X](https://savrn.com/ai-index/pricing/gpus/mi325x) $2.00 · [1x MI355X](https://savrn.com/ai-index/pricing/gpus/mi355x) $2.59 | Memory is the weights at that precision plus 20% for the runtime and a short context; a long context needs more. Prices are the lowest on-demand hourly rates in the [SAVRN Index](https://savrn.com/ai-index/pricing/gpus), read Oct 24, 2026. [MiniMax-M2.7 on every accelerator the SAVRN Index prices, at every precision](https://savrn.com/models/minimax-m2-7/gpus) ## SAVRN's Notes on MiniMax-M2.7 A 228.7B-parameter mixture-of-experts model where the cost line is the story. Working memory is 137.2 GB at 4-bit on one MI300X at $1.85 an hour, 274.4 GB at 8-bit on one MI355X at $2.59, and 548.9 GB at 16-bit on two MI355X at $5.18. Moving from 4-bit to 16-bit is a 2.8x jump in hourly cost, so settle your quantization before you reserve hardware — and only 8 of 256 experts fire per token, with FP8 in the published config and a 204,800-token context. If you would rather not own the hardware, two hosts on our Index price it per token: DeepInfra at $0.25 in and $1.00 out, Novita at $0.30 and $1.20 per million, both read Oct 1, 2026. Hosted wins below a few hundred million tokens a month and loses above it. The license is 'other' on our Index, so read the text before commercial use rather than assuming Apache. It draws 1.1M downloads a month, sitting next to DeepSeek-V4-Flash and MiMo. SAVRN Research ## Model Card Join Our WeChat Discord community. MiniMax Agent API CLI MiniMax Website Hugging Face GitHub ModelScope LICENSE MiniMax-M2.7 is our first model deeply participating in its own evolution. M2.7 is capable of building complex agent harnesses and completing highly elaborate productivity tasks, leveraging Agent Teams, complex Skills, and dynamic tool search. For more details, see our blog post. M2.7 initiates a cycle of model self-evolution: during development, we let the model update its own memory, build dozens of complex skills for RL experiments, and improve its own learning process based on experiment results. An internal version of M2.7 autonomously optimized a programming scaffold over… Excerpt from the card by MiniMax, licensed other. ## Configuration Architecture MiniMaxM2ForCausalLM Context length (tokens) 204,800 Layers 62 Hidden size 3,072 Feed-forward size 1,536 Attention heads 48 Key/value heads 8 Head dimension 128 Vocabulary size 200,064 Experts 256 Experts active per token 8 RoPE base 5,000,000 Model type minimax_m2 Quantization fp8 ## Identity and Version Repository MiniMaxAI/MiniMax-M2.7 Publisher MiniMax Task Text generation Modality Text Library transformers Parameters 228.7B parameters Languages Not stated by the source Revision d494266a4affc0d2995ba1fa35c8481cbd84294b First published 2026-04-09 Last updated 2026-04-20 ## Files and Weights 151 files, 230.2 GB in total. The weights are 125 files totalling 230.1 GB in safetensors. Weights125 files · 230.1 GB Configuration5 files · 10.3 MB Tokenizer4 files · 16.9 MB Documentation10 files · 74.3 KB Other6 files · 8.1 MB Repository1 file · 1.8 KB Every file | File | Type | Size | SHA-256 | | --- | --- | --- | --- | | model-00000-of-00130.safetensors | Weights | 3.7 GB | 9785f5a87c85 | | model-00001-of-00130.safetensors | Weights | 1.2 GB | d2ed94efe077 | | model-00002-of-00130.safetensors | Weights | 2.5 GB | f0c1b97aff37 | | model-00003-of-00130.safetensors | Weights | 1.2 GB | 93be479ff1b6 | | model-00004-of-00130.safetensors | Weights | 2.5 GB | 5d5bead700b8 | | model-00005-of-00130.safetensors | Weights | 1.2 GB | 99444d6d83c6 | | model-00006-of-00130.safetensors | Weights | 2.5 GB | df42d1d91b84 | | model-00007-of-00130.safetensors | Weights | 1.2 GB | 18882ffcb4f2 | | model-00008-of-00130.safetensors | Weights | 2.5 GB | cf8ead5d7b01 | | model-00009-of-00130.safetensors | Weights | 1.2 GB | d897820ce912 | | model-00010-of-00130.safetensors | Weights | 2.5 GB | 734eee6e6286 | | model-00011-of-00130.safetensors | Weights | 1.2 GB | 1237cbe1b991 | | model-00012-of-00130.safetensors | Weights | 2.5 GB | 069b272af352 | | model-00013-of-00130.safetensors | Weights | 1.2 GB | 045403b45c89 | | model-00014-of-00130.safetensors | Weights | 2.5 GB | 0277da3d1063 | | model-00015-of-00130.safetensors | Weights | 1.2 GB | d2a9db97dbab | | model-00016-of-00130.safetensors | Weights | 2.5 GB | 90776eaf1438 | | model-00017-of-00130.safetensors | Weights | 1.2 GB | 4ea50b70dae5 | | model-00018-of-00130.safetensors | Weights | 2.5 GB | 2a239e9eae27 | | model-00019-of-00130.safetensors | Weights | 1.2 GB | 5e041358d2ce | | model-00020-of-00130.safetensors | Weights | 2.5 GB | 4f4f7af9ded3 | | model-00021-of-00130.safetensors | Weights | 1.2 GB | 8a76ddac0582 | | model-00022-of-00130.safetensors | Weights | 2.5 GB | c080ad8c3b50 | | model-00023-of-00130.safetensors | Weights | 1.2 GB | 9eee017222d3 | | model-00024-of-00130.safetensors | Weights | 2.5 GB | e3d3c543000e | | model-00025-of-00130.safetensors | Weights | 1.2 GB | 68580bdb4da6 | | model-00026-of-00130.safetensors | Weights | 2.5 GB | c0ca69318b53 | | model-00027-of-00130.safetensors | Weights | 1.2 GB | a6f03ff04b01 | | model-00028-of-00130.safetensors | Weights | 2.5 GB | 6432450282a2 | | model-00029-of-00130.safetensors | Weights | 1.2 GB | 961ca8675f7e | | model-00030-of-00130.safetensors | Weights | 2.5 GB | 7687ab86a251 | | model-00031-of-00130.safetensors | Weights | 1.2 GB | 345042a45204 | | model-00032-of-00130.safetensors | Weights | 2.5 GB | 4faa680a93c4 | | model-00033-of-00130.safetensors | Weights | 1.2 GB | fdfa10d9c831 | | model-00034-of-00130.safetensors | Weights | 2.5 GB | ae23de77bccd | | model-00035-of-00130.safetensors | Weights | 1.2 GB | 6a5ca9a1fd87 | | model-00036-of-00130.safetensors | Weights | 2.5 GB | 88113822767b | | model-00037-of-00130.safetensors | Weights | 1.2 GB | 3a42e3dfe02d | | model-00038-of-00130.safetensors | Weights | 2.5 GB | 6cf2b3485504 | | model-00039-of-00130.safetensors | Weights | 1.2 GB | bbf5e9eff764 | | model-00040-of-00130.safetensors | Weights | 2.5 GB | 499c9039dff0 | | model-00041-of-00130.safetensors | Weights | 1.2 GB | 3ed0565052bb | | model-00042-of-00130.safetensors | Weights | 2.5 GB | 601959ff7bdb | | model-00043-of-00130.safetensors | Weights | 1.2 GB | 7fbd3484ee80 | | model-00044-of-00130.safetensors | Weights | 2.5 GB | b349ca4c4779 | | model-00045-of-00130.safetensors | Weights | 1.2 GB | 54673ecdf05e | | model-00046-of-00130.safetensors | Weights | 2.5 GB | 341ac0c20e20 | | model-00047-of-00130.safetensors | Weights | 1.2 GB | 38785114c81c | | model-00048-of-00130.safetensors | Weights | 2.5 GB | 59c01cf8b22f | | model-00049-of-00130.safetensors | Weights | 1.2 GB | bbc2141546a2 | | model-00050-of-00130.safetensors | Weights | 2.5 GB | e6c1dfceca02 | | model-00051-of-00130.safetensors | Weights | 1.2 GB | bc4209a8554b | | model-00052-of-00130.safetensors | Weights | 2.5 GB | e1ad313b24dc | | model-00053-of-00130.safetensors | Weights | 1.2 GB | 84f5bb1d8a74 | | model-00054-of-00130.safetensors | Weights | 2.5 GB | a001ec5d2dd1 | | model-00055-of-00130.safetensors | Weights | 1.2 GB | da2a90dda71a | | model-00056-of-00130.safetensors | Weights | 2.5 GB | 9fe32d8911b7 | | model-00057-of-00130.safetensors | Weights | 1.2 GB | 1e8d73847187 | | model-00058-of-00130.safetensors | Weights | 2.5 GB | 61ae96c27243 | | model-00059-of-00130.safetensors | Weights | 1.2 GB | 2f19cb6bc24a | | model-00060-of-00130.safetensors | Weights | 2.5 GB | 3b43a6164a76 | | model-00061-of-00130.safetensors | Weights | 1.2 GB | 0ed80e6f71a5 | | model-00062-of-00130.safetensors | Weights | 2.5 GB | b89ff1d3c45e | | model-00063-of-00130.safetensors | Weights | 1.2 GB | 8585c7cd9418 | | model-00064-of-00130.safetensors | Weights | 2.5 GB | 593a3e7a56cf | | model-00065-of-00130.safetensors | Weights | 1.2 GB | 75849a0106d8 | | model-00066-of-00130.safetensors | Weights | 2.5 GB | 00408d159353 | | model-00067-of-00130.safetensors | Weights | 1.2 GB | 4bface08c504 | | model-00068-of-00130.safetensors | Weights | 2.5 GB | 4a168b285a43 | | model-00069-of-00130.safetensors | Weights | 1.2 GB | 246239f37d0a | | model-00070-of-00130.safetensors | Weights | 2.5 GB | 080f36a819d8 | | model-00071-of-00130.safetensors | Weights | 1.2 GB | 9e1a3b6d59ca | | model-00072-of-00130.safetensors | Weights | 2.5 GB | 3702d9c9f31f | | model-00073-of-00130.safetensors | Weights | 1.2 GB | c71864e0febd | | model-00074-of-00130.safetensors | Weights | 2.5 GB | 07e6d2b9d5cf | | model-00075-of-00130.safetensors | Weights | 1.2 GB | 674d2be3b866 | | model-00076-of-00130.safetensors | Weights | 2.5 GB | 160a131a07cb | | model-00077-of-00130.safetensors | Weights | 1.2 GB | 2a2a1eee70e8 | | model-00078-of-00130.safetensors | Weights | 2.5 GB | 799eaaf53b6f | | model-00079-of-00130.safetensors | Weights | 1.2 GB | 5bf243b40049 | | model-00080-of-00130.safetensors | Weights | 2.5 GB | 9f97809043ca | | model-00081-of-00130.safetensors | Weights | 1.2 GB | 8129bb648b2b | | model-00082-of-00130.safetensors | Weights | 2.5 GB | 4a581d6de6af | | model-00083-of-00130.safetensors | Weights | 1.2 GB | 246f02a0e291 | | model-00084-of-00130.safetensors | Weights | 2.5 GB | 4083ad0a522b | | model-00085-of-00130.safetensors | Weights | 1.2 GB | 35ea447ad683 | | model-00086-of-00130.safetensors | Weights | 2.5 GB | 257545b54e89 | | model-00087-of-00130.safetensors | Weights | 1.2 GB | 0f9a088db132 | | model-00088-of-00130.safetensors | Weights | 2.5 GB | 55ec4e69a22d | | model-00089-of-00130.safetensors | Weights | 1.2 GB | e86e9b192f49 | | model-00090-of-00130.safetensors | Weights | 2.5 GB | b57125eec75a | | model-00091-of-00130.safetensors | Weights | 1.2 GB | 0c613cdacd62 | | model-00092-of-00130.safetensors | Weights | 2.5 GB | a406fcc45a8a | | model-00093-of-00130.safetensors | Weights | 1.2 GB | 27f4f5084a43 | | model-00094-of-00130.safetensors | Weights | 2.5 GB | d6db2523f161 | | model-00095-of-00130.safetensors | Weights | 1.2 GB | a8480d9cc921 | | model-00096-of-00130.safetensors | Weights | 2.5 GB | 0123cfef652f | | model-00097-of-00130.safetensors | Weights | 1.2 GB | 181466337b86 | | model-00098-of-00130.safetensors | Weights | 2.5 GB | cb371f55564e | | model-00099-of-00130.safetensors | Weights | 1.2 GB | 9f0f0bd9e07f | | model-00100-of-00130.safetensors | Weights | 2.5 GB | 45fd433c26aa | | model-00101-of-00130.safetensors | Weights | 1.2 GB | 82ed509a2950 | | model-00102-of-00130.safetensors | Weights | 2.5 GB | af2c3743f403 | | model-00103-of-00130.safetensors | Weights | 1.2 GB | 4ac162b3348b | | model-00104-of-00130.safetensors | Weights | 2.5 GB | 2f21dabd4f42 | | model-00105-of-00130.safetensors | Weights | 1.2 GB | 97dd9fc182eb | | model-00106-of-00130.safetensors | Weights | 2.5 GB | 55f519426d24 | | model-00107-of-00130.safetensors | Weights | 1.2 GB | b50498a9bf40 | | model-00108-of-00130.safetensors | Weights | 2.5 GB | 502bdd08025b | | model-00109-of-00130.safetensors | Weights | 1.2 GB | 5a528b71b52c | | model-00110-of-00130.safetensors | Weights | 2.5 GB | 20d1fa5b1659 | | model-00111-of-00130.safetensors | Weights | 1.2 GB | f63b6c84659c | | model-00112-of-00130.safetensors | Weights | 2.5 GB | c4e0b5428019 | | model-00113-of-00130.safetensors | Weights | 1.2 GB | e48bfe3f2a38 | | model-00114-of-00130.safetensors | Weights | 2.5 GB | 08fb5a9fd032 | | model-00115-of-00130.safetensors | Weights | 1.2 GB | adf4ab941b45 | | model-00116-of-00130.safetensors | Weights | 2.5 GB | 45ba476d7a50 | | model-00117-of-00130.safetensors | Weights | 1.2 GB | 7f2a7438c4f6 | | model-00118-of-00130.safetensors | Weights | 2.5 GB | a2eac3b06b70 | | model-00119-of-00130.safetensors | Weights | 1.2 GB | 5cf0b16b764d | | model-00120-of-00130.safetensors | Weights | 2.5 GB | 16a1f3697a69 | | model-00121-of-00130.safetensors | Weights | 1.2 GB | b2eabbce0590 | | model-00122-of-00130.safetensors | Weights | 2.5 GB | 2d4060a5e532 | | model-00123-of-00130.safetensors | Weights | 1.2 GB | 2779fd92da6e | | model-00124-of-00130.safetensors | Weights | 1.2 GB | 3439acf43dfe | | config.json | Configuration | 1.7 KB | — | | configuration_minimax_m2.py | Configuration | 10.2 KB | — | | generation_config.json | Configuration | 166 B | — | | model.safetensors.index.json | Configuration | 10.2 MB | — | | modeling_minimax_m2.py | Configuration | 30.9 KB | — | | LICENSE | Documentation | 3.7 KB | — | | README.md | Documentation | 14.9 KB | — | | docs/sglang_deploy_guide.md | Documentation | 3.8 KB | — | | docs/sglang_deploy_guide_cn.md | Documentation | 3.7 KB | — | | docs/tool_calling_guide.md | Documentation | 16.6 KB | — | | docs/tool_calling_guide_cn.md | Documentation | 16.7 KB | — | | docs/transformers_deploy_guide.md | Documentation | 3.0 KB | — | | docs/transformers_deploy_guide_cn.md | Documentation | 2.9 KB | — | | docs/vllm_deploy_guide.md | Documentation | 4.6 KB | — | | docs/vllm_deploy_guide_cn.md | Documentation | 4.4 KB | — | | chat_template.jinja | Other | 6.5 KB | — | | figures/agent_harness.png | Other | 311.9 KB | 7c661c39ff84 | | figures/agent_teams.gif | Other | 7.4 MB | 2f72c28868b6 | | figures/banner.png | Other | 119.6 KB | d524f03ea8db | | figures/benchmark_overview.png | Other | 78.4 KB | — | | figures/mle_bench.png | Other | 122.7 KB | 6bdbff9fc7f9 | | .gitattributes | Repository | 1.8 KB | — | | merges.txt | Tokenizer | 2.4 MB | — | | tokenizer.json | Tokenizer | 9.7 MB | — | | tokenizer_config.json | Tokenizer | 10.9 KB | — | | vocab.json | Tokenizer | 4.7 MB | — | ## License and Download License other Access Open weights, no gate Download size 230.1 GB [Download from MiniMax](https://huggingface.co/MiniMaxAI/MiniMax-M2.7) Released by MiniMax through its official repository on Hugging Face. [Read the license](https://github.com/MiniMax-AI/MiniMax-M2.7/blob/main/LICENSE). ## Evaluations Each result is shown as reported, with the conditions its reporter stated. None is a SAVRN measurement. A comparison lines two results up only when their configuration, unit and setup are all stated and identical. | Benchmark | Conditions | Result | Reported by | Revision | Date | | --- | --- | --- | --- | --- | --- | | SWE-bench/SWE-bench_Multilingual | Task swe_bench_multilingual_%_resolvedMetric swe_bench_multilingual_%_resolvedComparison conditions not established | 76.5 | [Model Card](https://huggingface.co/MiniMaxAI/MiniMax-M2.7) Reported by a third party | Evaluated revision not stated | 2026-08-10 | | ScaleAI/SWE-bench_Pro | Task SWE_Bench_ProMetric SWE_Bench_ProComparison conditions not established | 56.2 | [Model Card](https://huggingface.co/MiniMaxAI/MiniMax-M2.7) Reported by a third party | Evaluated revision not stated | 2026-04-12 | | benchflow/skillsbench | Task skillsbench_v1_1Metric skillsbench_v1_1Setup with-skills; BenchFlow harness; OpenHands agent; 87 tasks x 3 trials; full 261/261 coverageComparison conditions not established | 34.9 | [SkillsBench v1.1 official leaderboard](https://huggingface.co/datasets/benchflow/skillsbench-leaderboard/raw/main/leaderboard/skillsbench/v1.1/official.json) Reported by a third party | Evaluated revision not stated | 2026-06-11 | | claw-eval/Claw-Eval | Task generalMetric generalSetup Pass³% \| N=3 \| 161 tasksComparison conditions not established | 49.7 | Claw-Eval Leaderboard Reported by a third party | Evaluated revision not stated | 2026-04-23 | | claw-eval/Claw-Eval | Task multi_turnMetric multi_turnSetup Pass³% \| N=3 \| 38 tasksComparison conditions not established | 44.7 | Claw-Eval Leaderboard Reported by a third party | Evaluated revision not stated | 2026-04-23 | | harborframework/terminal-bench-2.0 | Task terminalbench_2Metric terminalbench_2Comparison conditions not established | 57 | [Model Card](https://huggingface.co/MiniMaxAI/MiniMax-M2.7) Reported by a third party | Evaluated revision not stated | 2026-04-12 | | internlm/WildClawBench | Task avg_costMetric avg_costComparison conditions not established | 7.2 | WildClawBench Reported by a third party | Evaluated revision not stated | 2026-05-22 | | internlm/WildClawBench | Task avg_timeMetric avg_timeComparison conditions not established | 551 | WildClawBench Reported by a third party | Evaluated revision not stated | 2026-05-22 | | internlm/WildClawBench | Task overallMetric overallComparison conditions not established | 33.8 | WildClawBench Reported by a third party | Evaluated revision not stated | 2026-05-22 | ## Memory Requirements | Precision | Weights in memory | | --- | --- | | As published | 230.1 GB | | 16-bit | 457.4 GB | | 8-bit | 228.7 GB | | 4-bit | 114.3 GB | Weights only, from the published parameter count; the key-value cache and runtime add to this. ## Hosted Prices | Host | Input / output | Unit | Observed | | --- | --- | --- | --- | | DeepInfra | $0.25 / $1.00 | input / output, per million tokens | Oct 24, 2026 | | Novita | $0.30 / $1.20 | input / output, per million tokens | Oct 24, 2026 | From the [SAVRN Index](https://savrn.com/ai-index/pricing/open-models). ## Compare MiniMax-M2.7 - [DeepSeek-V4-Flash-0731 vs MiniMax-M2.7](https://savrn.com/models/compare/deepseek-v4-flash-0731-vs-minimax-m2-7) ## Questions About MiniMax-M2.7 ### How much GPU memory does MiniMax-M2.7 need? About 548.9 GB at 16-bit and 137.2 GB at 4-bit: the weights (228.7B parameters) plus a working margin. A long context needs more. ### What is the cheapest GPU to run MiniMax-M2.7 on? At 16-bit, 2x MI355X from $5.18 an hour; at 4-bit, 1x MI300X from $1.85 an hour, at the lowest on-demand prices the SAVRN Index lists. ### What license is MiniMax-M2.7 released under? other, as its publisher declares it. Read the license text before commercial use. ### What is MiniMax-M2.7's context length? 204,800 tokens, from the maximum position embeddings in its published configuration. ## Similar Models Model · Text generation ### [Hy3-Razor-226B-A18B-E144of192](https://savrn.com/models/hy3-razor-226b-a18b-e144of192) [Mingyang Song](https://savrn.com/model-publishers/nickyang) Hy3 with a quarter of its routed experts removed by RAZOR, a training-free expert pruning method. Every MoE layer keeps 144 of its original 192 routed experts. No gradient updates or recovery training were applied: the retained weights are the base model's own weights. Pruning touches only the routed expert pool. Attention, shared experts, the embedding and the LM head are untouched, so the compute per token drops only by the share of expert FLOPs that the removed experts would have contributed. The other budget is Requires a Transformers build containing the native hyv3 implementation. Weights are bfloat16. RAZOR asks whether the surviving computation can replace an expert's function… Open weights apache-2.0 226.3B parameters 262,144 tokens transformers [View model](https://savrn.com/models/hy3-razor-226b-a18b-e144of192) Model · Text generation ### [Qwen3-235B-A22B-Thinking-2507](https://savrn.com/models/qwen3-235b-a22b-thinking-2507) [Zachary Howard](https://savrn.com/model-publishers/servantofares) Over the past three months, we have continued to scale the thinking capability of Qwen3-235B-A22B, improving both the quality and depth of reasoning. We are pleased to introduce Qwen3-235B-A22B-Thinking-2507, featuring the following key enhancements: - Significantly improved performance on reasoning tasks, including logical reasoning, mathematics, science, coding, and academic benchmarks that typically require human expertise — achieving state-of-the-art results among open-source thinking models. - Markedly better general capabilities, such as instruction following, tool usage, text generation, and alignment with human preferences. - Enhanced 256K long-context understanding capabilities.… Open weights apache-2.0 235.1B parameters 262,144 tokens transformers [View model](https://savrn.com/models/qwen3-235b-a22b-thinking-2507) Model · Text generation ### [DeepSeek-V4.1-Flash-EXL3-3bpw-2x-RTX-PRO-6000](https://savrn.com/models/deepseek-v4-1-flash-exl3-3bpw-2x-rtx-pro-6000) [Diffbot](https://savrn.com/model-publishers/diffbot) DeepSeek-V4.1-Flash, ready to serve on two RTX PRO 6000 Blackwell cards (sm120, 96 GB each) with vLLM at tensor-parallel 2. The routed experts are - int4 Engram tables; - a 3-bit DSpark drafter; - a serving stack that keeps 30% of the routed experts (the ones agentic coding uses least) in pinned host RAM and runs them next to the VRAM experts; - a decode-once prefill kernel for the 3-bit experts. The recipe/ folder has the image build, the sm120 patches, the MoE kernel, and the serving, benchmark, KL and quantization scripts. We measured KL against the original model with uses the original model's top-512 logprobs on three corpora, 49,152 scored positions each. The neutral corpus is… Open weights mit 219B parameters 1,048,576 tokens [View model](https://savrn.com/models/deepseek-v4-1-flash-exl3-3bpw-2x-rtx-pro-6000) Model · Text generation ### [Qwen3.8-Flash-Next-oQ4e-mix6-mtp](https://savrn.com/models/qwen3-8-flash-next-oq4e-mix6-mtp) [Umixer](https://savrn.com/model-publishers/umixer) Goal: oQ4e speed and memory footprint, oQ5e-or-better quality reaching toward oQ6e. Runs at ~73–75 GB resident with ~82→66 tok/s decode on a 128 GB Apple Silicon host — the speed/RAM economics of a 4-bit build, at an accuracy tier equal to or above uniform oQ5e. How: a mixed-precision MLX build of Qwen3.8-Flash-Next, assembled by component-wise byte-copy from Jundot's native oQ quantizations — zero re-quantization in the entire package. Every tensor is byte-identical to its source, so quality is traceable, not re-derived. Uniform-bit quantization pays accuracy where it isn't needed (middle layers) and loses it where it hurts (first/last layers, n-gram table, attention). A Data-Free… Open weights other 180B parameters 262,144 tokens mlx [View model](https://savrn.com/models/qwen3-8-flash-next-oq4e-mix6-mtp) Model · Text generation ### [DeepSeek-V4-Flash](https://savrn.com/models/deepseek-v4-flash) [DeepSeek](https://savrn.com/model-publishers/deepseek-ai) We present a preview version of DeepSeek-V4 series, including two strong Mixture-of-Experts (MoE) language models — DeepSeek-V4-Pro with 1.6T parameters (49B activated) and DeepSeek-V4-Flash with 284B parameters (13B activated) — both supporting a context length of one million tokens. DeepSeek-V4 series incorporate several key upgrades in architecture and optimization: 1. Hybrid Attention Architecture: We design a hybrid attention mechanism combining Compressed Sparse Attention (CSA) and Heavily Compressed Attention (HCA) to dramatically improve long-context efficiency. In the 1M-token context setting, DeepSeek-V4-Pro requires only 27% of single-token inference FLOPs and 10% of KV cache… Open weights mit 290.9B parameters 1,048,576 tokens transformers [View model](https://savrn.com/models/deepseek-v4-flash) Model · Text generation ### [DeepSeek-V4-Flash-DSpark](https://savrn.com/models/deepseek-v4-flash-dspark) [DeepSeek](https://savrn.com/model-publishers/deepseek-ai) Note: DeepSeek-V4-Flash-DSpark is not a new model. It is the same checkpoint with an additional speculative decoding module attached. A minimal inference example is available in the inference folder. For more details, refer to: https://github.com/deepseek-ai/DeepSpec We present a preview version of DeepSeek-V4 series, including two strong Mixture-of-Experts (MoE) language models — DeepSeek-V4-Pro with 1.6T parameters (49B activated) and DeepSeek-V4-Flash with 284B parameters (13B activated) — both supporting a context length of one million tokens. DeepSeek-V4 series incorporate several key upgrades in architecture and optimization: 1. Hybrid Attention Architecture: We design a hybrid… Open weights mit 165.3B parameters 1,048,576 tokens transformers [View model](https://savrn.com/models/deepseek-v4-flash-dspark) ## MiniMax [All models and datasets](https://savrn.com/model-publishers/minimaxai) ## Versions - [d494266a4aff](https://savrn.com/models/minimax-m2-7/versions/d494266a4aff) · current 2026-10-06 ## Explore More - [All text generation models](https://savrn.com/models/tasks/text-generation) - [Model comparisons](https://savrn.com/models/comparisons) - [The model directory](https://savrn.com/models) - [Open model prices by host](https://savrn.com/ai-index/pricing/open-models) ## Source - Repository metadata, read 2026-10-19. - [MiniMax website](https://www.minimax.io/) - [Hugging Face record](https://huggingface.co/MiniMaxAI/MiniMax-M2.7) - [How the hub is built](https://savrn.com/model-hub/methodology)