Model Follower.

Up to date

Rebuilt
2026-10-26 07:36 UTC
Release radar
116 on watch
Hub snapshot
2026-10-26
Re-checked
daily · every number sourced and dated

Models › Kimi-K2 family

Kimi-K2-Instruct-0905

Hugging Face ↗ Follow changes

moonshotai Kimi (Modified MIT) (modified-mit) out 14 months ago

1.03T parameters (8 of 384 experts per token), text in, text out, 256K tokens of context. 1 provider serves it from $0.60 per million input tokens.

Verified 2026-10-26 07:27 UTC Rank #152 in LLMs · score 92.3 method v1 Sources 4 Downloads 30d 27.7k · #430 of 619 tracked

Can I run it?616GB at INT4Estimateneeds a multi-GPU nodemethod
What does it cost?$0.60/1M incheapest of 1Novitafp82026-10-26
Can I use it?Kimi (Modified MIT) (modified-mit)MIT plus a display-the-model-name attribution requirement for products above 100M MAU or US$20M monthly revenue; the exact wording is in each repo's LICENSElicence file2026-10-26
Do people use it?27.7kdownloads, 30 days#430 of 619 we track · below the median+0 likes, downloads −45% this week790 likesHugging Face2026-10-26

Worth reading

Picked by rule from 19 tracked pieces: the lab's own post, the most shared, a field report, the latest.

Most discussed · 936 HN pts

Kimi K2 Thinking, a SOTA open-source trillion-parameter reasoning model

Kimi K2 Thinking is an open-source thinking agent model that executes up to 300 sequential tool calls to solve complex problems.

  • The model marks latest efforts in test-time scaling by scaling thinking tokens and tool calling steps.
  • K2 Thinking is live on kimi.com under the chat mode with full agentic mode available soon.
moonshotai.github.io 2025-11-06 936 HN pts · 427 comments 34k words about the Kimi K2 family family
Latest · 4 months ago

Kimi K2.7 Code is generally available in GitHub Copilot

Kimi K2.7 Code is generally available in GitHub Copilot as the first open-weight model offered as a selectable option in the Copilot model picker.

github.blog 2026-07-02 417 HN pts · 185 comments 317 words about the Kimi K2 family family

19 best of 19 tracked

The 2 picks above, plus these 17. Top is the radar's score: discussion, source authority and the lab's own posts, fading with age.

  1. ArticleKimi K2.6: Advancing open-source codingkimi.com · 2026-04-20 · 710 HN pts · official
  2. ArticleKimi Released Kimi K2.5, Open-Source Visual SOTA-Agentic Modelkimi.com · 2026-01-27 · 502 HN pts · official
  3. DiscussionKimi K2 is a state-of-the-art mixture-of-experts (MoE) language modeltwitter.com · 2025-07-11 · 348 HN pts
  4. DiscussionKimi K2 1T model runs on 2 512GB M3 Ultrastwitter.com · 2025-12-14
  5. CodeKimi K2 is a state-of-the-art mixture-of-experts (MoE) language modelgithub.com · 2025-07-12 · 352 HN pts
  6. CodeKimi-K2 Tech Report [pdf]github.com · 2025-07-21 · 84 HN pts
  7. ArticleKimi-K2-Instruct-0905 - Simon Willison's Weblogsimonwillison.net · undated
  8. ArticleKimi K2 0905 - Intelligence, Performance & Price Analysisartificialanalysis.ai · undated
  9. ArticleComparing GPT-5.4, Opus 4.6, GLM-5.1, Kimi K2.5, MiMo V2 Pro and MiniMax M2.7codejam.info · 2026-04-18 · 6 HN pts
  10. ArticleMoonshot AI's Kimi K2 outperforms GPT-4 in key benchmarks – and it's freeventurebeat.com · 2025-07-13
  11. DiscussionKimi-K2-Thinking 1T at 28 tok/SEC with MLX tensor paralleltwitter.com · 2025-12-19
  12. ArticleKimi K2-Instruct-0905: Specs, Benchmarks, Pricing | AI/TLDRai-tldr.dev · undated
  13. ArticleKimi K2 Instruct vs Kimi K2-Instruct-0905: Benchmarks, Pricing & Which Is Better in 2026llm-stats.com · undated
  14. ArticleKimi K2–0905 Review - Mediummedium.com · undated
  15. ArticleKimi K2 0905 Benchmarks, Pricing & Context Window - LLM Statsllm-stats.com · undated
  16. Articlehttps://intuitionlabs.ai/articles/kimi-k2-technical-deep-diveintuitionlabs.ai · undated
  17. ArticleKimi K2-0905: 1T Open MoE Built for Agents & Codingdigitalapplied.com · 2025-09-04

Where to run it

Live hosted endpoints, cheapest first, USD per million tokens. Ties break on output price, then provider name.

#Cheapest live rowIn /1MOut /1MContextQuantSeen
1 Novita $0.60 $2.50 256K fp8 Oct 26

All 1 provider →

Where it's used

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; 16 quantized builds for local use.

  • As a service1Novita
  • On own hardware16quantized builds
  • Built on64 fine-tunes · 2 adapters · 0 merges
  • In apps100+Hugging Face Spaces

Top Spaces running it: enzostvs/deepsite ♥ 16,604 · hadadxyz/ai ♥ 819 · umint/ai ♥ 258 · dropkickJesus999/deepsiteV3 ♥ 32

Will it fit?

Weights only, an estimate: leave room for the KV cache. Method

Needs an 8×80 GB node even at INT4

About 616 GB at INT4, 1.23 TB at FP8, 2.46 TB at BF16: weights only, so leave room for the KV cache. Every expert must be resident, so total parameters count.

16 GB card64 GB Mac2×80 GB8×80 GBINT4: about 616 GB of weights (estimate; KV cache and activations extra)INT4616 GBFP8: about 1.23 TB of weights (estimate; KV cache and activations extra)FP81.23 TBBF16: about 2.46 TB of weights (estimate; KV cache and activations extra)BF162.46 TB
16 GB card64 GB Mac2×80 GB8×80 GBINT4: about 616 GB of weights (estimate; KV cache and activations extra)INT4616 GBFP8: about 1.23 TB of weights (estimate; KV cache and activations extra)FP81.23 TBBF16: about 2.46 TB of weights (estimate; KV cache and activations extra)BF162.46 TB

Estimate params × bytes per param × 1.2 · how we estimate

The story so far

Weights, every dated article, and repo commits on one rail.

8 dated articles and 17 repo commits so far

Teal is the weights (2025-09-03), amber triangles are 8 dated articles (each links to it), grey ticks are repo commits.

AugSepOctNovDecJanFebMarAprMayJunJulAugSepOct2026-01-30: 1 repo commit2025-11-07: 1 repo commit2025-10-22: 3 repo commits2025-10-10: 1 repo commit2025-09-05: 4 repo commits2025-09-04: 2 repo commits2025-09-03: 5 repo commitsKimi K2.7 Code is generally available in GitHub Copilot · 2026-07-02Kimi K2.6: Advancing open-source coding · 2026-04-20Kimi Released Kimi K2.5, Open-Source Visual SOTA-Agentic Model · 2026-01-27Kimi K2 Thinking, a SOTA open-source trillion-parameter reasoning model · 2025-11-06Kimi K2 is a state-of-the-art mixture-of-experts (MoE) language model · 2025-07-11Kimi K2 is a state-of-the-art mixture-of-experts (MoE) language model · 2025-07-12Kimi-K2 Tech Report [pdf] · 2025-07-21Comparing GPT-5.4, Opus 4.6, GLM-5.1, Kimi K2.5, MiMo V2 Pro and MiniMax M2.7 · 2026-04-18Weights 2025-09-03 · 2025-09-03Weights 2025-09-03now · 1 provider
AugSepOctNovDecJanFebMarAprMayJunJulAugSepOct2026-01-30: 1 repo commit2025-11-07: 1 repo commit2025-10-22: 3 repo commits2025-10-10: 1 repo commit2025-09-05: 4 repo commits2025-09-04: 2 repo commits2025-09-03: 5 repo commitsKimi K2.7 Code is generally available in GitHub Copilot · 2026-07-02Kimi K2.6: Advancing open-source coding · 2026-04-20Kimi Released Kimi K2.5, Open-Source Visual SOTA-Agentic Model · 2026-01-27Kimi K2 Thinking, a SOTA open-source trillion-parameter reasoning model · 2025-11-06Kimi K2 is a state-of-the-art mixture-of-experts (MoE) language model · 2025-07-11Kimi K2 is a state-of-the-art mixture-of-experts (MoE) language model · 2025-07-12Kimi-K2 Tech Report [pdf] · 2025-07-21Comparing GPT-5.4, Opus 4.6, GLM-5.1, Kimi K2.5, MiMo V2 Pro and MiniMax M2.7 · 2026-04-18Weights 2025-09-03 · 2025-09-03Weights 2025-09-03now · 1 provider

What people built on it

22 derivatives on the Hub. Which build to run, by runtime: the most downloaded in each class.

22 derivatives on the Hub, most of them quantized

From the Hugging Face model tree; quantized builds in teal, fine-tunes, merges and adapters in greys.

16 quantized (Hugging Face model tree)4 fine-tunes (Hugging Face model tree)2 adapters (Hugging Face model tree)16 quantized4 fine-tunes2 adapters
16 quantized (Hugging Face model tree)4 fine-tunes (Hugging Face model tree)2 adapters (Hugging Face model tree)16 quantized4 fine-tunes2 adapters
GGUF unsloth/Kimi-K2-Instruct-0905-GGUF for llama.cpp, Ollama, LM Studio 61.8k downloads MLX mlx-community/Kimi-K2-Instruct-0905-mlx-DQ3_K_M for Apple silicon 19.8k downloads

All variants →

Reference

Everything else, folded: the full tables, every source.

Benchmarks

1 row

Benchmarks 1 row

measured third-party leaderboard rowauthor-reported number from the model card or eval files (1)No confidence intervals unless the source gave them; settings differ between rows, so compare within a benchmark, not across. Date = the benchmark's own date when it gives one, else when we read it.
BenchmarkValueSetting / notesReported byDate
author-reported SWE-bench Verified · Write code with agents69.2 ± 0.63—model author2026-10-07

Use-case evidence

4 use cases

What it is good at 4 use cases

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.

Facts

21 fields

Listed in LLMs · MoE

Facts

Pipeline tagtext-generation
Librarytransformers
ArchitecturesDeepseekV3ForCausalLM
Model typekimi_k2
Dtypebfloat16
Params by dtypeBF16: 2,514,975,360 · F32: 23,040 · F8_E4M3: 1,023,893,241,856
Size bandXXL
MoEyes
Experts384
Experts per token8
Context length262,144 (config.json max_position_embeddings)
LicenseKimi (Modified MIT) (modified-mit) (other)
Hub snapshot2026-10-26

Servability

vLLM · SGLang

Servability

vLLM: supported · DeepseekV3ForCausalLM · 4b1c04e76fe9ac3ca8b7e541dd423e7d5f38dc60SGLang: supported · DeepseekV3ForCausalLM · 60d941293e3af27a12c6eb64e2c3e2e1a17f100f

vLLM registry (4b1c04e76fe9ac3ca8b7e541dd423e7d5f38dc60) · SGLang registry (60d941293e3af27a12c6eb64e2c3e2e1a17f100f) · Matched by architecture class name against each project's registry on main; 'supported' means the architecture is registered, not that this checkpoint was tested.

Hardware fit

3 estimates

Hardware fit ESTIMATE

Weights-only memory for Kimi-K2-Instruct-0905: params × bytes/param × 1.2 overhead. KV cache and activations are extra. How this is computed.

BF16 (2 bytes/param)2.46 TB · Cluster (≥640 GB)
FP8 (1 byte/param)1.23 TB · Cluster (≥640 GB)
INT4 (0.5 bytes/param)616 GB · Multi-GPU node (<640 GB)

Weights only (params × bytes/param × 1.2 overhead); KV cache and activations extra. MoE: all experts must be resident, so total parameters count.

All variants

22 variants

Lineage

Original release (shipped quantized by the lab) Family: moonshotai/Kimi-K2 22 derivatives on the Hub

quantized: 16adapter: 2finetune: 4

Top variants (22 shown, ranked by downloads)

Relation
Quant
22 rows
Variant Relation Quant Params Band License Author Likes Downloads
unsloth/Kimi-K2-Instruct-0905-GGUF quantized GGUF — — other unsloth 53 61.8k
mlx-community/Kimi-K2-Instruct-0905-mlx-DQ3_K_M quantized mlx 1.03T XXL other mlx-community 12 19.8k
bartowski/moonshotai_Kimi-K2-Instruct-0905-GGUF quantized GGUF — — — bartowski 2 16.6k
ubergarm/Kimi-K2-Instruct-0905-GGUF quantized GGUF — — other ubergarm 16 12.8k
DevQuasar/moonshotai.Kimi-K2-Instruct-0905-GGUF quantized GGUF — — — DevQuasar 1 6.1k
amd/Kimi-K2-Instruct-0905-MXFP4 quantized MXFP4 550B XXL other amd 1 5.8k
gghfez/Kimi-K2-Instruct-0905-control-vectors adapter GGUF — — cc-by-4.0 gghfez 0 4.6k
richardyoung/Kimi-K2-Instruct-0905-MLX-2bit quantized MLX 1.03T XXL apache-2.0 richardyoung 2 3.8k
mlx-community/Kimi-K2-Instruct-0905-mlx-3bit quantized mlx 1.03T XXL other mlx-community 3 3.1k
richardyoung/Kimi-K2-Instruct-0905-MLX-8bit quantized MLX 1.03T XXL apache-2.0 richardyoung 0 3k
unsloth/Kimi-K2-Instruct-0905-BF16 finetune -BF16 1.03T XXL other unsloth 4 2.3k
richardyoung/Kimi-K2-Instruct-0905-MLX-4bit quantized MLX 1.03T XXL apache-2.0 richardyoung 1 2.4k
richardyoung/Kimi-K2-Instruct-0905-MLX-3bit quantized MLX 1.03T XXL apache-2.0 richardyoung 0 2.4k
richardyoung/Kimi-K2-Instruct-0905-MLX-6bit quantized MLX 1.03T XXL apache-2.0 richardyoung 0 2.1k
noctrex/Kimi-K2-Instruct-0905-MXFP4_MOE-GGUF finetune GGUF — — — noctrex 0 2k
unsloth/Kimi-K2-Instruct-0905 quantized — 1.03T XXL other unsloth 7 1.3k
anikifoss/Kimi-K2-Instruct-0905-HQ4_K quantized GGUF — — other anikifoss 3 1.1k
gghfez/Kimi-K2-0905-no-not-X-but-Y-cv adapter GGUF — — cc-by-4.0 gghfez 3 1k
DevQuasar/moonshotai.Kimi-K2-Instruct-0905-BF16 quantized -BF16 1.03T XXL — DevQuasar 0 914
mradermacher/Kimi-K2-Instruct-0905-i1-GGUF quantized GGUF — — other mradermacher 2 472
mradermacher/Kimi-K2-Instruct-0905-GGUF finetune GGUF — — other mradermacher 0 0
dasLOL/Kimi-K2-Instruct-0905 finetune — — — — dasLOL 0 0

Repository changelog

17 commits

Repository changelog 17 commits

DateCommitSHA
2026-01-30 Transformers v5 support (#21) ac6c49f
2025-11-07 fix-vocab-size (#17) 12d9c7c
2025-10-22 not add functions. in tool id (#15) 94a4053
2025-10-22 remove-auto-add-str-functions (#14) 975af05
2025-10-22 fix apply_chat_template (#13) 46ec167
2025-10-10 update_chat_template_and_tokenizer (#12) 09d5f93
2025-09-05 Update README.md 7152993
2025-09-05 Update README.md 2ce3235
2025-09-05 Update README.md e4f9671
2025-09-05 Update README.md 65191cf
2025-09-04 Update README.md d30fdf6
2025-09-04 add FAQ for tool calls. d56abb2
2025-09-03 update readme 13adc1d
2025-09-03 Add files using upload-large-folder tool b1ad81b
2025-09-03 Add files using upload-large-folder tool eb1e816
2025-09-03 Upload config.json with huggingface_hub 440d282
2025-09-03 initial commit 86c87f9

All providers

1 provider · 1 endpoint

Where to run 1 providers · 1 endpoints

Hosted endpoints observed for Kimi-K2-Instruct-0905, sorted by input price (USD per million tokens). Each row names its observer and date; prices change daily, so check the source before quoting. Cheapest observed input price: $0.60/M.

ProviderModel refContextQuant$/M in$/M outtok/sTTFTToolsStatusObserved
Novita
hf, openrouter
 moonshotai/kimi-k2-0905 256K fp8 $0.60 $2.50 78.9 1775 ms yes live 2026-10-26

Articles and coverage

19 pieces

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.

Official2

Kimi K2.6: Advancing open-source coding
kimi.com2026-04-20officialauthorityHN front pageabout the Kimi K2 familyHN 710 pts · 372 comments2,299 words
Kimi K2.6 is an open-source model released for advanced coding, long-horizon execution, and agent swarm capabilities. It is available via Kimi.com, the Kimi App, the API, and Kimi Code, showing a 12% increase in code generation accuracy and a 96.60% tool invocation success rate. The model scales agent swarms to 300 sub-agents executing across 4,000 coordinated steps simultaneously.written, reviewed, cited
  • Kimi K2.6 optimizes model inference in Zig, a highly niche programming language, to demonstrate exceptional out-of-distribution generalization.
  • The model reconfigured the exchange-core thread topology from 4ME+2RE to 2ME+1RE to extract a 185% medium throughput leap.
Kimi Released Kimi K2.5, Open-Source Visual SOTA-Agentic Model
kimi.com2026-01-27officialauthorityHN front pageabout the Kimi K2 familyHN 502 pts · 239 comments1,856 words
Kimi K2.5 self-directs an agent swarm with up to 100 sub-agents executing parallel workflows across up to 1,500 tool calls. This swarm architecture reduces execution time by up to 4.5x compared to a single-agent setup without predefined workflows.generated, cited

Analysis7

Kimi K2.7 Code is generally available in GitHub Copilot
github.blog2026-07-02authorityHN front pageabout the Kimi K2 familyHN 417 pts · 185 comments317 words
Kimi K2.7 Code is generally available in GitHub Copilot as the first open-weight model offered as a selectable option in the Copilot model picker. The model is beginning to roll out to Copilot Pro, Pro+, and Max plans within Visual Studio Code version 1.127.0 or later, while it remains off by default for Copilot Business and Copilot Enterprise requiring plan administrators to enable the policy. Rollout will be gradual as GitHub continues to monitor the model's quality and performance.written, reviewed, cited
Kimi K2 Thinking, a SOTA open-source trillion-parameter reasoning model
moonshotai.github.io2025-11-06HN front pageabout the Kimi K2 familyHN 936 pts · 427 comments34,041 words
Kimi K2 Thinking is an open-source thinking agent model that executes up to 300 sequential tool calls to solve complex problems. The model achieves 44.9% on HLE with tools, 60.2% on BrowseComp, and 71.3% on SWE-Bench Verified. It demonstrated deep reasoning by solving a PhD-level mathematics problem through 23 interleaved reasoning and tool calls.written, reviewed, cited
  • The model marks latest efforts in test-time scaling by scaling thinking tokens and tool calling steps.
  • K2 Thinking is live on kimi.com under the chat mode with full agentic mode available soon.
Kimi K2 is a state-of-the-art mixture-of-experts (MoE) language model
github.com2025-07-12authorityHN front pageabout the Kimi K2 familyHN 352 pts · 2 comments2,355 words
Kimi-K2 Tech Report [pdf]
github.com2025-07-21authorityabout the Kimi K2 familyHN 84 pts · 6 comments
Kimi-K2-Instruct-0905 - Simon Willison's Weblog
simonwillison.netauthority203 words
Kimi K2 0905 - Intelligence, Performance & Price Analysis
artificialanalysis.aiauthorityabout the Kimi K2 family964 words
Comparing GPT-5.4, Opus 4.6, GLM-5.1, Kimi K2.5, MiMo V2 Pro and MiniMax M2.7
codejam.info2026-04-18about the Kimi K2 familyHN 6 pts · 2 comments1,457 words
GPT-5.4 achieved an average cleanliness score of 9 while GLM-5.1 scored 8 in the technical approach evaluation. Kimi K2.5 and MiniMax M2.7 were consistently ranked at the bottom with average scores of 5.6 and 5.2 respectively.generated, cited

Community1

Kimi K2 is a state-of-the-art mixture-of-experts (MoE) language model
twitter.com2025-07-11HN front pageabout the Kimi K2 familyHN 348 pts · 179 comments
All coverage tracked for this model (19) — append-only, nothing is ever removed
TitleSourcePublishedFirst seenSignals
Kimi-K2-Instruct-0905 - Simon Willison's Weblogsimonwillison.net—2026-10-07exa
Kimi K2 0905 - Intelligence, Performance & Price Analysisartificialanalysis.ai—2026-10-07exa
Kimi K2-Instruct-0905: Specs, Benchmarks, Pricing | AI/TLDRai-tldr.dev—2026-10-07exa
Kimi K2 Instruct vs Kimi K2-Instruct-0905: Benchmarks, Pricing & Which Is Better in 2026llm-stats.com—2026-10-07exa
Kimi K2–0905 Review - Mediummedium.com—2026-10-07exa
Kimi K2 0905 Benchmarks, Pricing & Context Window - LLM Statsllm-stats.com—2026-10-07exa
https://intuitionlabs.ai/articles/kimi-k2-technical-deep-diveintuitionlabs.ai—2026-10-07exa
Kimi K2.7 Code is generally available in GitHub Copilotgithub.blog2026-07-022026-10-07hn
Kimi K2.6: Advancing open-source codingkimi.com official2026-04-202026-10-07hn
Comparing GPT-5.4, Opus 4.6, GLM-5.1, Kimi K2.5, MiMo V2 Pro and MiniMax M2.7codejam.info2026-04-182026-10-07hn
Kimi Released Kimi K2.5, Open-Source Visual SOTA-Agentic Modelkimi.com official2026-01-272026-10-07hn
Kimi-K2-Thinking 1T at 28 tok/SEC with MLX tensor paralleltwitter.com2025-12-192026-10-07hn
Kimi K2 1T model runs on 2 512GB M3 Ultrastwitter.com2025-12-142026-10-07hn
Kimi K2 Thinking, a SOTA open-source trillion-parameter reasoning modelmoonshotai.github.io2025-11-062026-10-07hn
Kimi K2-0905: 1T Open MoE Built for Agents & Codingdigitalapplied.com2025-09-042026-10-07exa
Kimi-K2 Tech Report [pdf]github.com2025-07-212026-10-07hn
Moonshot AI's Kimi K2 outperforms GPT-4 in key benchmarks – and it's freeventurebeat.com2025-07-132026-10-07hn
Kimi K2 is a state-of-the-art mixture-of-experts (MoE) language modelgithub.com2025-07-122026-10-07hn
Kimi K2 is a state-of-the-art mixture-of-experts (MoE) language modeltwitter.com2025-07-112026-10-07hn

Similar models

3 models

Similar models

  • MiMo-V2.6-Pro-RL · 1.02T
    same size band (XXL), also MoE, same task (text-generation), fewer providers, permissive license
  • MiMo-V2.5-Pro · 1.02T
    same size band (XXL), also MoE, same task (text-generation), fewer providers, permissive license
  • GLM-5.3 · 753B
    same size band (XXL), also MoE, same task (text-generation), more providers

Sources

4 sources

Sources

Hub snapshot: 2026-10-26 · HF API: 2026-10-26 07:16 UTC · Providers: 2026-10-26 07:26 UTC · Articles: 2026-10-26 07:27 UTC · Summary: 2026-10-07 15:01 UTC · Last verified 2026-10-26 07:27 UTC

Follow Kimi-K2-Instruct-0905

Watch this model

Email alerts are not open yet; the feeds above are the same data. See feeds for the item schema.