The story so far
Weights, every dated article, and repo commits on one rail.
12 repo commits since the weights landed
Teal is the weights (2025-02-22), grey ticks are repo commits.
Moonlight-16B-A3B-Instruct
moonshotai MIT out 20 months ago
16.0B parameters, 3B active (6 of 64 experts per token), text in, text out, 8K tokens of context.
Verified 2026-10-25 Rank #621 in LLMs · score 73.6 method v1 Sources 3 Downloads 30d 34.5k · #413 of 619 tracked
Do people use it?34.5kdownloads, 30 days#413 of 619 we track · below the median+1 likes, downloads +3% this week209 likesHugging Face2026-10-24
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.
No API provider serves it, so its downloads are people running it themselves: 9 quantized builds and 5 Spaces.
Will it fit?
Weights only, an estimate: leave room for the KV cache. Method
Fits a 16 GB card at INT4
About 9.6 GB at INT4, 19 GB at FP8, 38 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
What people built on it
12 derivatives on the Hub. Which build to run, by runtime: the most downloaded in each class.
12 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.
Reference
Everything else, folded: the full tables, every source.
Benchmarks
11 rows
Benchmarks 11 rows
measured third-party leaderboard rowauthor-reported number from the model card or eval files (11)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.
| Benchmark | Value | Setting / notes | Reported by | Date |
|---|---|---|---|---|
| author-reported BBH · Reason through hard problems | 65.2 percent | — | model author | 2026-10-07 |
| author-reported C-Eval · Work across languages | 77.2 percent | — | model author | 2026-10-07 |
| author-reported CMMLU · Work across languages | 78.2 percent | — | model author | 2026-10-07 |
| author-reported Commonsense QA sets · Chat and follow instructions | 66.3 percent | — | model author | 2026-10-07 |
| author-reported GSM8K · Reason through hard problems | 77.4 percent | — | model author | 2026-10-07 |
| author-reported HumanEval · Write code with agents | 48.1 percent | — | model author | 2026-10-07 |
| author-reported MATH-500 · Reason through hard problems | 45.3 percent | — | model author | 2026-10-07 |
| author-reported MATH-500 · Reason through hard problems | 81.1 percent | — | model author | 2026-10-07 |
| author-reported MBPP · Write code with agents | 63.8 percent | — | model author | 2026-10-07 |
| author-reported MMLU · Reason through hard problems | 70 percent | — | model author | 2026-10-07 |
| author-reported MMLU-Pro · Reason through hard problems | 42.4 percent | — | model author | 2026-10-07 |
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.
- Chat and follow instructionsauthor-reportedAuthor reports TriviaQA‡ 66.3decision-model estimate p=1.00name: Instruct+2 more
- Reason through hard problemsauthor-reportedAuthor reports MMLU 70.0Author reports MMLU-pro 42.4Author reports BBH 65.2+2 more
- Write code with agentsauthor-reportedAuthor reports HumanEval 48.1Author reports MBPP 63.8
- Work across languagesauthor-reportedAuthor reports C-Eval 77.2Author reports CMMLU 78.2
Facts
23 fields
Facts
- Moonlight-16B-A3B-Instruct has 16.0B parameters (3B active per token). (safetensors metadata, observed 2026-10-24)
- Moonlight-16B-A3B-Instruct is a mixture-of-experts model with 64 experts, routing 6 per token. (config.json, observed 2026-10-24)
- Context length is 8,192 tokens (8K). (config.json max_position_embeddings, observed 2026-10-24)
- License: MIT (friction: none). Permissive. (license file, observed 2026-10-24)
- Created 2025-02-22, last modified 2026-01-30. (HF API, observed 2026-10-24)
| Pipeline tag | text-generation |
|---|---|
| Library | transformers |
| Architectures | DeepseekV3ForCausalLM |
| Model type | deepseek_v3 |
| Dtype | bfloat16 |
| Params by dtype | BF16: 15,960,111,936 |
| Size band | L |
| MoE | yes |
| Experts | 64 |
| Experts per token | 6 |
| Context length | 8,192 (config.json max_position_embeddings) |
| Papers | arXiv:2502.16982 |
| License | MIT (mit) |
| Hub snapshot | 2026-10-25 |
Servability
vLLM · SGLang
Servability
vLLM: supported · DeepseekV3ForCausalLM · 7050ef881ff0835f6890a47db1859a24c65481d6SGLang: supported · DeepseekV3ForCausalLM · 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.
Hardware fit
3 estimates
Hardware fit ESTIMATE
Weights-only memory for Moonlight-16B-A3B-Instruct: params × bytes/param × 1.2 overhead. KV cache and activations are extra. How this is computed.
| BF16 (2 bytes/param) | 38 GB · Single high-end GPU (<80 GB) |
|---|---|
| FP8 (1 byte/param) | 19 GB · Single GPU (<24 GB) |
| INT4 (0.5 bytes/param) | 9.6 GB · Single GPU (<24 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
12 variants
Lineage
quantized: 9finetune: 3
Top variants (12 shown, ranked by downloads)
| Variant | Relation | Quant | Params | Band | License | Author | Likes | Downloads |
|---|---|---|---|---|---|---|---|---|
| mmnga/Moonlight-16B-A3B-Instruct-gguf | quantized | gguf | — | — | mit | mmnga | 14 | 85.4k |
| gabriellarson/Moonlight-16B-A3B-Instruct-GGUF | quantized | GGUF | — | — | mit | gabriellarson | 7 | 16.6k |
| AbteeXAILab/lumynax-moe-moonlight-16b-a3b-gguf | quantized | gguf | — | — | mit | AbteeXAILab | 0 | 2.9k |
| mlx-community/Moonlight-16B-A3B-Instruct-4-bit | quantized | — | 16.0B | L | mit | mlx-community | 4 | 2.2k |
| Myric/Moonlight-16B-A3B-Instruct-APEX-GGUF | quantized | GGUF | — | — | mit | Myric | 0 | 1.5k |
| noctrex/Moonlight-16B-A3B-Instruct-MXFP4_MOE-GGUF | quantized | GGUF | — | — | — | noctrex | 3 | 1.1k |
| slowfastai/Moonlight-16B-A3B-Instruct-bnb-4bit | quantized | -bnb | 16.4B | L | mit | slowfastai | 2 | 664 |
| rawcell/Moonlight-16B-A3B-Instruct-bruno | finetune | — | 16.0B | L | apache-2.0 | rawcell | 1 | 325 |
| sasa2000/Moonlight-16B-A3B-Instruct-int4-AutoRound | quantized | int4 | 675M | S | mit | sasa2000 | 0 | 327 |
| rawcell/Moonlight-16B-A3B-Instruct-abliterated | finetune | — | 16.0B | L | apache-2.0 | rawcell | 1 | 137 |
| sasa2000/Moonlight-16B-A3B-Instruct-FP8-Dynamic | quantized | -FP8 | 16.0B | L | mit | sasa2000 | 0 | 74 |
| twilson63/Moonlight-16B-A3B-int4-ov | finetune | int4 | — | — | mit | twilson63 | 0 | 0 |
Repository changelog
12 commits
Repository changelog 12 commits
| Date | Commit | SHA |
|---|---|---|
| 2026-01-30 | Transformers v5 support (#16) | 4e735b0 |
| 2025-03-03 | hotfix cache_utils max_cache_length (#10) | 336ed0d |
| 2025-02-26 | Update README.md | 930af5a |
| 2025-02-26 | Update modeling_deepseek.py (#4) | fc14f0a |
| 2025-02-22 | Update README.md | 9558325 |
| 2025-02-22 | update readme | b42671f |
| 2025-02-22 | update readme | 579a3a3 |
| 2025-02-22 | Update README.md | 4308c42 |
| 2025-02-22 | update readme format | d413755 |
| 2025-02-22 | Update README.md | bb39c64 |
| 2025-02-22 | first commit | 391e7a8 |
| 2025-02-22 | initial commit | 5883873 |
Papers
1 paper
Papers 1
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).
- Muon is Scalable for LLM Training cited in card
Similar models
3 models
Similar models
- DeepSeek-Coder-V2-Lite-Instruct · 15.7B same size band (L), also MoE, same task (text-generation)
- gpt-oss-20b · 20.9B same size band (L), also MoE, same task (text-generation), more providers, permissive license
- DeepSeek-V2-Lite · 15.7B same size band (L), also MoE, same task (text-generation)
Sources
3 sources
Sources
Hub snapshot: 2026-10-25 · HF API: 2026-10-24 21:41 UTC · Providers: 2026-10-24 21:44 UTC · Articles: 2026-10-24 21:46 UTC · Summary: 2026-10-07 15:13 UTC · Last verified 2026-10-25
Follow Moonlight-16B-A3B-Instruct
Watch this model
- /feeds/models/moonshotai--Moonlight-16B-A3B-Instruct.xml(RSS 2.0) · every change to Moonlight-16B-A3B-Instruct
- /feeds/families/moonshotai--Moonlight.xml(RSS 2.0) · every change across the Moonlight family
Email alerts are not open yet; the feeds above are the same data. See feeds for the item schema.