Up to date
moonshotai MIT out 12 months ago
49.1B parameters, 3B active, text in, text out. 1 provider serves it.
Verified 2026-10-25 Rank #252 in LLMs · score 85.8 method v1 Sources 3 Downloads 30d 143.4k · #277 of 619 tracked
Changed this week: 1 new quantized build on the Hub (25 in all).2026-10-24
Picked by rule from 12 tracked pieces: the lab's own post, the most shared, a field report, the latest.
Kimi Linear is a hybrid linear attention architecture that outperforms full attention across short, long, and reinforcement learning contexts.
Kimi Linear 48B Instruct emerges as the best generalist model, while Qwen3 Coder Next serves as an immediate replacement for coding tasks on a homelab featuring 2× AMD AI Max+ 395 GPUs with 128GB unified memory each.
The 3 picks above, plus these 9. Top is the radar's score: discussion, source authority and the lab's own posts, fading with age.
Live hosted endpoints, cheapest first, USD per million tokens. Ties break on output price, then provider name.
| # | Cheapest live row | In /1M | Out /1M | Context | Quant | Seen |
|---|---|---|---|---|---|---|
| 1 | Featherless | — | — | — | — | Oct 25 |
Hugging Face does not say who downloads a model or with what. These are the proxies: API providers listing it, quantized builds, fine-tunes/adapters/merges, and Spaces that load it.
Served by 1 API provider; 25 quantized builds for local use.
Weights only, an estimate: leave room for the KV cache. Method
About 30 GB at INT4, 59 GB at FP8, 118 GB at BF16: weights only, so leave room for the KV cache. Every expert must be resident, so total parameters count.
Estimate params × bytes per param × 1.2 · how we estimate
Weights, every dated article, and repo commits on one rail.
Teal is the weights (2025-10-30), amber triangles are 4 dated articles (each links to it), grey ticks are repo commits.
38 derivatives on the Hub. Which build to run, by runtime: the most downloaded in each class.
From the Hugging Face model tree; quantized builds in teal, fine-tunes, merges and adapters in greys.
Everything else, folded: the full tables, every source.
| Benchmark | Value | Setting / notes | Reported by | Date |
|---|---|---|---|---|
| author-reported MMLU-Pro · Reason through hard problems | 51 | — | model author | 2026-10-07 |
| author-reported RULER · Read long documents | 84.3 | — | model author | 2026-10-07 |
Labels are evidence, not verdicts: measured a third-party leaderboard row for this model; author-reported a number the model author published in the card or eval files; structural a deterministic signal from the name, config or chat template, or a decision-model estimate. Each chip links to its source. All use cases.
| Pipeline tag | text-generation |
|---|---|
| Library | transformers |
| Architectures | KimiLinearForCausalLM |
| Model type | kimi_linear |
| Dtype | bfloat16 |
| Params by dtype | BF16: 49,122,599,168 · F32: 82,560 |
| Size band | XL |
| MoE | yes |
| Experts | 256 |
| Papers | arXiv:2510.26692, arXiv:2412.06464 |
| License | MIT (mit) |
| Hub snapshot | 2026-10-25 |
vLLM: supported · KimiLinearForCausalLM · 7050ef881ff0835f6890a47db1859a24c65481d6SGLang: supported · KimiLinearForCausalLM · 773cb938fe9484a8f7e7e46fb1b1fe96412c59ed
vLLM registry (7050ef881ff0835f6890a47db1859a24c65481d6) · SGLang registry (773cb938fe9484a8f7e7e46fb1b1fe96412c59ed) · Matched by architecture class name against each project's registry on main; 'supported' means the architecture is registered, not that this checkpoint was tested.
Weights-only memory for Kimi-Linear-48B-A3B-Instruct: params × bytes/param × 1.2 overhead. KV cache and activations are extra. How this is computed.
| BF16 (2 bytes/param) | 118 GB · Multi-GPU node (<640 GB) |
|---|---|
| FP8 (1 byte/param) | 59 GB · Single high-end GPU (<80 GB) |
| INT4 (0.5 bytes/param) | 30 GB · Single high-end GPU (<80 GB) |
Weights only (params × bytes/param × 1.2 overhead); KV cache and activations extra. MoE: all experts must be resident, so total parameters count.
finetune: 10quantized: 25adapter: 3
| Variant | Relation | Quant | Params | Band | License | Author | Likes | Downloads |
|---|---|---|---|---|---|---|---|---|
| yujiepan/kimi-linear-tiny-random | finetune | — | 5.65M | S | — | yujiepan | 0 | 462.1k |
| cyankiwi/Kimi-Linear-48B-A3B-Instruct-AWQ-4bit | quantized | AWQ | 9.37B | L | mit | cyankiwi | 18 | 122.3k |
| bartowski/moonshotai_Kimi-Linear-48B-A3B-Instruct-GGUF | quantized | GGUF | — | — | mit | bartowski | 30 | 96.4k |
| mradermacher/Kimi-Linear-48B-A3B-Instruct-i1-GGUF | quantized | GGUF | — | — | mit | mradermacher | 3 | 29.4k |
| ymcki/Kimi-Linear-48B-A3B-Instruct-GGUF | quantized | GGUF | — | — | mit | ymcki | 37 | 23.8k |
| DhruvalLabs/Kimi-Linear-48B-A3B-Instruct-GGUF | quantized | GGUF | — | — | mit | DhruvalLabs | 3 | 17.3k |
| cerebras/Kimi-Linear-REAP-35B-A3B-Instruct | finetune | — | 35.1B | XL | mit | cerebras | 70 | 9.7k |
| mlx-community/Kimi-Linear-48B-A3B-Instruct-4bit | quantized | -4bit | 49.1B | XL | mit | mlx-community | 11 | 11.6k |
| Myric/Kimi-Linear-48B-A3B-Instruct-APEX-GGUF | quantized | GGUF | — | — | mit | Myric | 0 | 12k |
| AaryanK/Kimi-Linear-48B-A3B-Instruct-GGUF | quantized | GGUF | — | — | mit | AaryanK | 13 | 8.5k |
| noctrex/Kimi-Linear-48B-A3B-Instruct-MXFP4_MOE-GGUF | quantized | GGUF | — | — | — | noctrex | 4 | 6.8k |
| Firworks/Kimi-Linear-48B-A3B-Instruct-nvfp4 | quantized | nvfp4 | 28.0B | L | mit | Firworks | 10 | 6.4k |
| cyankiwi/Kimi-Linear-48B-A3B-Instruct-AWQ-8bit | quantized | AWQ | 15.3B | L | mit | cyankiwi | 4 | 6.4k |
| mradermacher/Kimi-Linear-48B-A3B-Instruct-GGUF | quantized | GGUF | — | — | mit | mradermacher | 1 | 4.1k |
| mlx-community/Kimi-Linear-48B-A3B-Instruct-6bit | quantized | — | 49.1B | XL | mit | mlx-community | 3 | 3.5k |
| mlx-community/Kimi-Linear-48B-A3B-Instruct-8bit | quantized | -8bit | 49.1B | XL | mit | mlx-community | 1 | 2.9k |
| brabooObrabo/Kimi-Linear-48B-A3B-Instruct-MXFP4-GS32-MLX | quantized | MXFP4 | 49.1B | XL | mit | brabooObrabo | 2 | 2.5k |
| mlx-community/Kimi-Linear-48B-A3B-Instruct-3bit | quantized | — | 49.1B | XL | mit | mlx-community | 2 | 1.6k |
| infosave/Kimi-Linear-48B-A3B-Code-CMF | quantized | — | — | — | mit | infosave | 0 | 1.4k |
| mlx-community/Kimi-Linear-48B-A3B-Instruct-mlx-bf16 | finetune | mlx | 49.1B | XL | mit | mlx-community | 1 | 1.1k |
| mlx-community/Kimi-Linear-48B-A3B-Instruct-5bit | quantized | — | 49.1B | XL | mit | mlx-community | 1 | 1.1k |
| JasmineBBB/Kimi-Linear-48B-A3B-Instruct-bnb-4bit | quantized | -bnb | 49.1B | XL | mit | JasmineBBB | 1 | 716 |
| mark-arts/kimi-linear-mlx-4bit | quantized | mlx | 49.1B | XL | mit | mark-arts | 0 | 733 |
| Kamali-Lab/Kimi-Linear-48B-A3B-Instruct-GGUF | quantized | GGUF | — | — | mit | Kamali-Lab | 2 | 605 |
| huawei-csl/Kimi-Linear-48B-A3B-Instruct-3bit-SINQ | quantized | — | 7.20B | M | apache-2.0 | huawei-csl | 1 | 416 |
| tiny-random/kimi-linear | finetune | — | 5.65M | S | — | tiny-random | 0 | 451 |
| Alpaczyk/Kimi-Linear-48B-A3B-Instruct-nvfp4 | quantized | nvfp4 | 28.0B | L | mit | Alpaczyk | 1 | 336 |
| huawei-csl/Kimi-Linear-48B-A3B-Instruct-4bit-SINQ | quantized | -4bit | 26.5B | L | apache-2.0 | huawei-csl | 3 | 202 |
| TribeBlend/tribeblend-etl-kimi-linear-48b-a3b-instruct | quantized | GGUF | — | — | other | TribeBlend | 0 | 260 |
| Optivise/kimi-linear-48b-a3b-instruct-qlora-fine-tuned | adapter | — | — | — | — | Optivise | 1 | 129 |
| Moonlight556/Kimi-Linear-48B-A3B-DFlash-240k | finetune | — | 451M | S | apache-2.0 | Moonlight556 | 0 | 124 |
| LeaderboardModel1/Kimi-Linear-48B-A3B-Instruct-AutoRound-W4A16-RTN | finetune | -W4A16 | 7.24B | M | — | LeaderboardModel1 | 0 | 65 |
| positron-ai/Kimi-Linear-48B-A3B-Instruct | finetune | — | 49.1B | XL | mit | positron-ai | 0 | 45 |
| masafy/kimi-fii-lora | adapter | — | — | — | cc-by-sa-3.0 | masafy | 0 | 22 |
| OistarsdataSolutions/NECRO-BOT | adapter | — | — | — | bigscience-bloom-rail-1.0 | OistarsdataSolutions | 0 | 0 |
| Rafael4882/dhx-teste | finetune | — | — | — | bigcode-openrail-m | Rafael4882 | 0 | 0 |
| jashansinghTT/kimi-linear-48b-a3b-instruct-blackhole | finetune | — | — | — | mit | jashansinghTT | 0 | 0 |
| mamingyuan/Kimi-Linear-48B-A3B-Instruct-DUET-v3 | finetune | — | — | — | other | mamingyuan | 0 | 0 |
| Date | Commit | SHA |
|---|---|---|
| 2025-12-16 | Fix deprecated import for Transformers v5 compatibility (#19) | e1df551 |
| 2025-11-25 | Update README.md | 806a95e |
| 2025-11-17 | Update README.md | 828c457 |
| 2025-11-17 | Update README.md | e6fa68d |
| 2025-11-17 | Upload arxiv.png | 9d37c46 |
| 2025-11-13 | Fix FLA import errors | d64a529 |
| 2025-11-01 | Update README.md | 6e163f3 |
| 2025-11-01 | Update README.md | 5eaec58 |
| 2025-11-01 | Update README.md | 06108b8 |
| 2025-11-01 | Update README.md | cbf1f86 |
| 2025-11-01 | Update README.md | e06dad3 |
| 2025-11-01 | Upload github.png | 3a7e0c5 |
| 2025-11-01 | Improve model card: Add metadata, paper, and GitHub links (#7) | cd777b7 |
| 2025-10-30 | Fix LaTeX rendering issue on README.md (#2) | fd1de63 |
| 2025-10-30 | Update README.md | 62a8341 |
| 2025-10-30 | Upload folder using huggingface_hub | 919416f |
| 2025-10-30 | initial commit | ff5066e |
Hosted endpoints observed for Kimi-Linear-48B-A3B-Instruct, sorted by input price (USD per million tokens). Each row names its observer and date; prices change daily, so check the source before quoting.
| Provider | Model ref | Context | Quant | $/M in | $/M out | tok/s | TTFT | Tools | Status | Observed |
|---|---|---|---|---|---|---|---|---|---|---|
| Featherless hf | moonshotai/Kimi-Linear-48B-A3B-Instruct | — | — | — | — | — | — | — | live | 2026-10-25 |
Ranked by source signals (HN, Lobsters, authority domains, official posts) and recency. Summaries are generated from a stored copy of the article text; every sentence cites a span of that text, and nothing is shown that failed the citation check.
arXiv links found in the model card and its tags. "Peer-reviewed" is shown only with evidence: a published venue in OpenAlex (measured) or the authors' own journal reference on arXiv (reported).
Hub snapshot: 2026-10-25 · HF API: 2026-10-24 21:36 UTC · Providers: 2026-10-24 21:44 UTC · Articles: 2026-10-24 21:45 UTC · Summary: 2026-10-07 15:04 UTC · Last verified 2026-10-25
Email alerts are not open yet; the feeds above are the same data. See feeds for the item schema.
