The story so far
Weights, every dated article, and repo commits on one rail.
15 repo commits since the weights landed
Teal is the weights (2024-03-28), grey ticks are repo commits.
Jamba-v0.1
ai21labs Apache 2.0 out 3 years ago
51.6B parameters (2 of 16 experts per token), text in, text out, 256K tokens of context.
Verified 2026-10-25 Rank #531 in LLMs · score 75.5 method v1 Sources 3 Downloads 30d 5.8k · #533 of 619 tracked
Do people use it?5.8kdownloads, 30 days#533 of 619 we track · below the median+3 likes, downloads −11% this week1,204 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: 5 quantized builds and 21 Spaces.
- As a service0no API provider listed
- On own hardware5quantized builds
- Built on75 fine-tunes · 2 adapters · 0 merges
- In apps21Hugging Face Spaces
Top Spaces running it: KBaba7/Quant ♥ 51 · FallnAI/Quantize-HF-Models ♥ 19 · bhaskartripathi/LLM_Quantization ♥ 16 · openfree/LLM_Quantization ♥ 16
Will it fit?
Weights only, an estimate: leave room for the KV cache. Method
Fits a 64 GB Mac (unified memory) at INT4
About 31 GB at INT4, 62 GB at FP8, 124 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, led by 5 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
8 rows
Benchmarks 8 rows
measured third-party leaderboard rowauthor-reported number from the model card or eval files (8)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 ARC Challenge · Chat and follow instructions | 64.4% percent | — | model author | 2026-10-07 |
| author-reported BBH · Reason through hard problems | 45.4% percent | — | model author | 2026-10-07 |
| author-reported Commonsense QA sets · Chat and follow instructions | 83.2% percent | — | model author | 2026-10-07 |
| author-reported GSM8K · Reason through hard problems | 59.9% percent | — | model author | 2026-10-07 |
| author-reported HellaSwag · Chat and follow instructions | 87.1% percent | — | model author | 2026-10-07 |
| author-reported MMLU · Reason through hard problems | 67.4% percent | — | model author | 2026-10-07 |
| author-reported TruthfulQA · Chat and follow instructions | 46.4% percent | — | model author | 2026-10-07 |
| author-reported Winogrande · Chat and follow instructions | 82.5% percent | — | model author | 2026-10-07 |
What the card says
1 cited sentence
Use-case evidence
3 use cases
What it is good at 3 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 HellaSwag 87.1%Author reports Arc Challenge 64.4%Author reports WinoGrande 82.5%+2 more
- Reason through hard problemsauthor-reportedAuthor reports MMLU 67.4%Author reports BBH 45.4%Author reports GSM8K (CoT) 59.9%
- Read long documentsstructuraldecision-model estimate p=1.00config: 262,144 context
Facts
22 fields
Facts
- Jamba-v0.1 has 51.6B parameters. (safetensors metadata, observed 2026-10-24)
- Jamba-v0.1 is a mixture-of-experts model with 16 experts, routing 2 per token. (config.json, observed 2026-10-24)
- Context length is 262,144 tokens (256K). (config.json max_position_embeddings, observed 2026-10-24)
- License: Apache 2.0 (friction: none). Permissive; patent grant. (license file, observed 2026-10-24)
- Created 2024-03-28, last modified 2024-09-11. (HF API, observed 2026-10-24)
| Pipeline tag | text-generation |
|---|---|
| Library | transformers |
| Architectures | JambaForCausalLM |
| Model type | jamba |
| Dtype | bfloat16 |
| Params by dtype | BF16: 51,566,653,312 · F32: 3,670,016 |
| Size band | XL |
| MoE | yes |
| Experts | 16 |
| Experts per token | 2 |
| Context length | 262,144 (config.json max_position_embeddings) |
| Papers | arXiv:2403.19887 |
| License | Apache 2.0 (apache-2.0) |
| Hub snapshot | 2026-10-25 |
Servability
vLLM · SGLang
Servability
vLLM: supported · JambaForCausalLM · 7050ef881ff0835f6890a47db1859a24c65481d6SGLang: not in registry · 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 Jamba-v0.1: params × bytes/param × 1.2 overhead. KV cache and activations are extra. How this is computed.
| BF16 (2 bytes/param) | 124 GB · Multi-GPU node (<640 GB) |
|---|---|
| FP8 (1 byte/param) | 62 GB · Single high-end GPU (<80 GB) |
| INT4 (0.5 bytes/param) | 31 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.
All variants
12 variants
Lineage
finetune: 5quantized: 5adapter: 2
Top variants (12 shown, ranked by downloads)
| Variant | Relation | Quant | Params | Band | License | Author | Likes | Downloads |
|---|---|---|---|---|---|---|---|---|
| Severian/Jamba-Nexus-4xMoE | finetune | — | 17.7B | L | mit | Severian | 10 | 3.1k |
| DevQuasar/ai21labs.Jamba-v0.1-GGUF | quantized | GGUF | — | — | — | DevQuasar | 1 | 3.5k |
| Severian/Jamba-UltraInteract-Instruct-1B | finetune | — | 1.02B | S | apache-2.0 | Severian | 0 | 3.2k |
| mlabonne/Jambatypus-v0.1 | quantized | — | — | — | apache-2.0 | mlabonne | 40 | 1k |
| lightblue/Jamba-v0.1-chat-multilingual | finetune | — | 51.6B | XL | apache-2.0 | lightblue | 24 | 989 |
| Severian/Jamba-Hercules | quantized | — | 52.3B | XL | apache-2.0 | Severian | 12 | 835 |
| jondurbin/bagel-jamba-v05 | finetune | — | 51.6B | XL | apache-2.0 | jondurbin | 7 | 426 |
| royleibov/Jamba-v0.1-ZipNN-Compressed | finetune | — | — | — | apache-2.0 | royleibov | 1 | 646 |
| Pclanglais/Jambert | quantized | — | 52.4B | XL | apache-2.0 | Pclanglais | 5 | 366 |
| mattshumer/Jamba-Chat | adapter | — | — | — | — | mattshumer | 8 | 156 |
| KnutJaegersberg/jamba-bagel-4bit | quantized | -4bit | 51.6B | XL | apache-2.0 | KnutJaegersberg | 2 | 382 |
| mlabonne/Jambalpaca-v0.1 | adapter | — | 51.6B | XL | apache-2.0 | mlabonne | 7 | 129 |
Repository changelog
15 commits
Repository changelog 15 commits
| Date | Commit | SHA |
|---|---|---|
| 2024-09-11 | Update README.md | 9efd115 |
| 2024-08-20 | Update LoRA fine-tune example - more target_modules, lower LR, bf16 (#49) | c16867f |
| 2024-05-06 | Move to in-library checkpoint (#43) | ce13f3f |
| 2024-04-01 | Update README.md | 8ee14c3 |
| 2024-03-31 | Fix small typo in JambaDecoder (#24) | 86c5df0 |
| 2024-03-29 | Update README.md | 62df05e |
| 2024-03-29 | Update README.md | 37a7297 |
| 2024-03-28 | Remove TGI tag (#8) | a5b97ac |
| 2024-03-28 | Update modeling_jamba.py - LoRA support in Mamba (#6) | 512e755 |
| 2024-03-28 | Update README.md | ca4e687 |
| 2024-03-28 | Update README.md | 4116eee |
| 2024-03-28 | Update README.md | a6b0673 |
| 2024-03-28 | Update README.md | a7db71f |
| 2024-03-28 | upload model | 905f68a |
| 2024-03-28 | initial commit | 0776c31 |
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).
Similar models
3 models
Similar models
- Kimi-Linear-48B-A3B-Instruct · 49.1B same size band (XL), also MoE, same task (text-generation), more providers, permissive license
- Phi-3.5-MoE-instruct · 41.9B same size band (XL), also MoE, same task (text-generation), permissive license
- GLM-4.7-Flash · 31.2B also MoE, same task (text-generation), more providers, permissive license
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 Jamba-v0.1
Watch this model
- /feeds/models/ai21labs--Jamba-v0.1.xml(RSS 2.0) · every change to Jamba-v0.1
- /feeds/families/ai21labs--Jamba-v0.1.xml(RSS 2.0) · every change across the Jamba-v0.1 family
Email alerts are not open yet; the feeds above are the same data. See feeds for the item schema.