
Nimble v3 Shipped Quietly: Bespoke Labs' 9B Decision Model Went Non-Commercial
- OrcaNEWOrcaCyber: Hy4 Flash Abliterated / Uncensored2026-11-18$0.07 / $0.28 per 1M tokens · 185 tok/s
- openaiNEWOpenAI: GPT-6.2 Astra2026-11-17$8.00 / $40.00 per 1M tokens · 76 tok/s
- openaiNEWOpenAI: GPT-6.2 Sol2026-11-17$2.00 / $10.00 per 1M tokens · 151 tok/s
- tencentNEWTencent: Hy4 Flash2026-11-16$0.02 / $0.07 per 1M tokens · 240 tok/s
- AlibabaNEWQwen: Qwen4 Max2026-11-16$1.50 / $4.50 per 1M tokens · 178 tok/s
- GoogleNEWGoogle: Gemini 4 Carbon2026-11-10$2.00 / $10.00 per 1M tokens · 94 tok/s
- AnthropicNEWAnthropic: Claude Fable 5.22026-11-09$8.00 / $40.00 per 1M tokens · 212 tok/s
- OrcaOrcaCyber: GLM 5.3 Abliterated / Uncensored2026-11-04$3.00 / $5.00 per 1M tokens · 99 tok/s
- OrcaOrcaCyber: Hy4 Abliterated / Uncensored2026-11-04$0.80 / $1.60 per 1M tokens · 131 tok/s
- OrcaOrcaCyber: Kimi K3.1 Pro Abliterated / Uncensored2026-11-04$2.00 / $5.00 per 1M tokens · 82 tok/s
- tencentTencent: Hy42026-11-04$0.83 / $2.50 per 1M tokens · 122 tok/s
- OrcaOrcaCyber: MiniMax M3.1 Flash Abliterated / Uncensored2026-10-29$0.80 / $1.60 per 1M tokens · 105 tok/s
- OrcaOrcaCyber: GLM 5.3 Flash Abliterated / Uncensored2026-10-27$0.80 / $1.60 per 1M tokens · 68 tok/s
- openaiOpenAI: GPT-6.1 Terra2026-10-26$1.00 / $5.00 per 1M tokens · 36 tok/s
- OrcaOrcaCyber: Qwen3.8 Flash Next Abliterated / Uncensored2026-10-26$0.80 / $1.60 per 1M tokens · 89 tok/s
- OrcaOrcaCyber: MiMo V2.6 Flash Abliterated / Uncensored2026-10-26$0.80 / $1.60 per 1M tokens · 39 tok/s
- MoonshotMoonshotAI: Kimi K3.1 Pro2026-10-2447Intelligence
- OrcaOrcaCyber: DeepSeek V4.1 Flash Abliterated / Uncensored2026-10-23$0.80 / $1.60 per 1M tokens · 222 tok/s
- minimaxMiniMax: M3.1 Flash2026-10-1929Intelligence
- OrcaOrca: OrcaCyber Ultra 1.02026-10-17$5.00 / $10.00 per 1M tokens · 104 tok/s
The weights for Nimble v3 — filed as Bespoke-Nimble-9B-v3 — were uploaded to Hugging Face on September 30, 2026, and thirteen days later Bespoke Labs still has not said a word about them in public. There is no post on the lab's own blog, no release note, no entry in the repository's pull-request history that mentions the version, and no line in the project README's changelog. What exists is a five-file repository, a companion evaluation dataset, a short model card, and a licence that changed. The same lab also parked a placeholder for Bespoke-Nimble-27B-v3 on October 9 — a repository whose only model file is a note explaining that the weights do not exist yet.
That is the whole of what is knowable from the vendor side, and it is worth separating from what is not. Everything below is sourced from the Hugging Face API, the raw model cards, the GitHub API, and the packaging metadata the model now carries; where a number is the vendor's own, it says so.
What is actually in the repository
Bespoke-Nimble-9B-v3 contains README.md, adapter_config.json, adapter_model.safetensors, a schema_config.json, and .gitattributes. That is the complete list. It is a PEFT LoRA adapter, not a standalone model: the card names Qwen/Qwen3.5-9B as the base, and the base is not duplicated in the repository.
The card itself is four short sections. It says the adapter takes "a context and typed questions (a choice among options, or yes or no)" and returns a probability for each option. It reports a Decision Index 0.2.1 score of 56.88. It states the licence. It carries a citation naming Shreyas Pimpalgaonkar, Sid Devic, Mrinal Kumar, Pratik Dalal and Mahesh Sathiamoorthy.
What is not in it matters just as much. There is no tokenizer, no prompt builder, no reference inference code, no checksum file, no provenance record and no context-length contract — all of which the family's first release, Bespoke-Nimble-9B, shipped alongside its own weights on September 18 as part of a twenty-one-file package. The v3 repository is an adapter checkpoint and a licence line. Anything you need to run it correctly, you now assemble yourself or take from a packager.

The licence is the part that will decide who can use it
The one sentence in the v3 card that changes a procurement conversation is this: Bespoke-Nimble-9B-v3 is released under CC BY-NC 4.0, which the card glosses as "it may not be used for commercial purposes."
Every previous Nimble checkpoint was Apache 2.0. Bespoke-Nimble-9B is Apache 2.0. Bespoke-Nimble-9B-v2, uploaded September 23, is Apache 2.0. The base model is Apache 2.0 and remains so — Bespoke's card says so explicitly, which is the honest thing to do and also the reason a company cannot route around the restriction by pointing at the base licence. A permissively licensed base does not loosen a non-commercial adapter, and the adapter is where the decision behaviour lives.
There is a second licence question that is worse, because it is ambiguous rather than merely restrictive. The GitHub repository bespokelabsai/nimble — the project home, currently sitting at 2,179 stars — reports no licence at all through GitHub's API. A repository with no licence file is not open source in any sense a lawyer will accept, and that is true whether or not the maintainers intended it. The lab's own README acknowledges rough edges in the model's first release; an unlicensed training and curation codebase is a different class of problem.
So the practical position for anyone planning a product today is: Nimble v3's weights are readable, downloadable and non-commercial; the code that produced them has no declared terms. If you need a commercially usable decision model on the Qwen3.5-9B base, this is not the one, and the alternatives are examined in Nimble v3 vs Kev and Nimble v3 vs Decider.
Three sets of numbers, three different instruments
The v3 card carries exactly one score: 56.88 on Decision Index 0.2.1. Bespoke Labs' companion dataset repository publishes the full run, the engine and reproduction instructions, which is more transparency than most vendors offer, and the number is still the publisher's own — measured with its own reference implementation of a third-party kit.
For the earlier checkpoint the family published a much broader table, and it is the only head-to-head material that exists against the model this category is defined by. On 324 held-out contrastive pairs, the September 18 model matched 90.12% of reference labels against 93.21% for TypeSafe's Jev 1.13.0, with the Qwen3.5-9B base at 66.36%. On 3,880 human-labelled decisions across thirteen public datasets the margin was 75.7% against Jev's 76.0% — a dead heat. Those figures belong to the first Nimble, not to v3, and the project labels them as its own measurements.
The third set is the one nobody at Bespoke Labs ran. The package that now ships the model for local use published a 231-case run of the JevBench original, easy and hard sets, on an Apple M5 Max with a 40-core GPU: 201 of 231 correct at 88 ms median latency for the unquantized MLX build, 202 of 231 at 176 ms for the unquantized GGUF build, and 198 of 231 at 168 ms for the Q4_K_M quantisation — with the caveat printed next to the table that quantisation changes scores and decisions. This is a third-party harness, but it is a packaging harness, not an independent audit, and the packagers say outright that JevBench and the Decision Index use different tasks and different scoring.

What Bespoke Labs has not said
Five things a reader will reasonably expect to find and will not:
• A context length. The v3 card states none. The packager's listing advertises 32,768 tokens, and the family's earlier release documents an 8,192-token contract with oversized prompts rejected rather than truncated. Those are different numbers describing different things, and nothing in Bespoke's own v3 material resolves which applies to the adapter itself.
• An option ceiling. Earlier Nimble checkpoints documented 1 to 255 choices per field — a deliberate expansion beyond the original 26, using extra single-token vocabulary slots. Whether v3 inherits that ceiling is not stated.
• A temperature. The family's history here is genuinely useful and worth knowing: the original checkpoint shipped with a fitted value near 2.18, the September 24 update moved the default to 1.0 with no separate fit, and intermediate releases carried their own settings. Applying the wrong one does not change which answer wins, but it changes every probability — which is the entire product. The v3 card is silent.
• Any statement of what changed. The repository's PR history jumps from "Support latest Nimble 9B release and checkpoint defaults" on September 24 to a README badge on October 13. There is no v3 change note, no data-lineage update, no training summary.
• Anything at all about the 27B. The Bespoke-Nimble-27B-v3 repository, created October 9, contains a licence declaration, a README and a file whose entire content is that the weights are not released yet. The README gives one number — 65.33 on the Decision Index 0.3 public suite — and says the model runs on Bespoke Labs' hosted API, with access by contacting the lab. No download, no date, no weights. It is a placeholder and should be read as one.
That is the shape of a quiet shipment: a checkpoint appears, the documentation does not, and the version number is the only indication that a decision was made.
What actually changed this week
The September 30 upload is not what makes this writable today. What makes it writable is that Nimble v3 became installable in the last few days. A local model runtime added nimble-v3 to its public library on October 13, merging the adapter into the Qwen3.5-9B base so that one pull command yields a runnable 9B decision model in six weight formats — two unquantized builds at 17.9 GB, an 9.3 GB MXFP8 MLX build, a 6.9 GB NVFP4 build, a 9.5 GB Q8_0 and a 5.6 GB Q4_K_M. At the moment of writing the listing showed 45 downloads and an "updated three hours ago" stamp, which is about as fresh as a model catalogue entry gets.
Two things follow. The first is that the access question has changed character: thirteen days ago, using Nimble v3 meant downloading a 9B base, merging a LoRA, and reconstructing a prompt format the repository no longer documents. Now it means pulling a package — with the licence restriction travelling along inside it, and with the packager's own note that it merged the adapter into the publisher-specified base rather than redistributing anything Bespoke Labs did not publish.
The second is that a first-run number now exists for a model that had none. A 231-case accuracy figure and a latency distribution from a harness that is not the vendor's is exactly the kind of evidence this category has been missing, and it is worth reading as a floor rather than a verdict: the same run put the quantized build 4 cases below the unquantized one, and the packagers warn that your thresholds need re-validating against your own examples.
Where this leaves routing: a decision model is a narrow component, and Nimble v3 answers only the questions you hand it — it cannot write, explain, or return an answer that is not one of your options. The generative half of a pipeline like that is the part OrcaRouter covers: 200+ models behind one OpenAI-compatible key at provider list price passed through with 0% markup, with automatic failover and a routing DSL you can express as code. OrcaRouter does not serve Nimble v3, and nothing here should be read as a claim that it does. The accurate framing is that a cheap, local decision head and a routed generative model behind it are two halves of one pattern, and if you are building the second half, that is us.

What to do with it, and what to watch
If you are evaluating decision models rather than shipping one, Nimble v3 is now the cheapest way into the category: a single pull, no GPU required beyond a laptop, a documented third-party accuracy figure, and a licence that is fine for research. If you are building a product, the licence answers the question before the benchmarks do, and the missing context contract and temperature mean you will be calibrating on your own data whether you intended to or not.
The three things that would change this assessment: a vendor announcement that clarifies the licence (CC BY-NC on a flagship checkpoint is unusual for a lab that has otherwise shipped Apache 2.0 and raised $40M to build for enterprises), a populated 27B repository, and a change note explaining what v3 actually did differently from v2. Until then, the honest summary is that a competent team uploaded a checkpoint, restricted its use, and let the packagers do the announcing.
Nimble v3 is an open-weights adapter you serve yourself; the generative half of the same pipeline is what OrcaRouter covers. 200+ models behind one OpenAI-compatible key OrcaRouter does not serve Nimble v3, and nothing in this article should be read as a claim that it does.
