
MAI Decision 1 vs Nimble v3: A Hosted Scorer Against a Licence You Can't Ship
- GoogleNEWGoogle: Gemini 4 Carbon2026-11-10$2.00 / $10.00 per 1M tokens · 90 tok/s
- AnthropicNEWAnthropic: Claude Fable 5.22026-11-09$8.00 / $40.00 per 1M tokens · 97 tok/s
- orcaNEWOrca: OrcaC2 1.02026-11-0711 tok/s
- OrcaNEWOrcaCyber: GLM 5.3 Abliterated / Uncensored2026-11-04$3.00 / $5.00 per 1M tokens · 87 tok/s
- OrcaNEWOrcaCyber: Hy4 Abliterated / Uncensored2026-11-04$0.80 / $1.60 per 1M tokens · 134 tok/s
- OrcaNEWOrcaCyber: Kimi K3.1 Pro Abliterated / Uncensored2026-11-04$4.00 / $20.00 per 1M tokens · 86 tok/s
- OrcaOrcaCyber: MiniMax M3.1 Flash Abliterated / Uncensored2026-10-29$0.80 / $1.60 per 1M tokens · 124 tok/s
- OrcaOrcaCyber: GLM 5.3 Flash Abliterated / Uncensored2026-10-27$0.80 / $1.60 per 1M tokens · 67 tok/s
- OrcaOrcaCyber: Qwen3.8 Flash Next Abliterated / Uncensored2026-10-26$0.80 / $1.60 per 1M tokens · 91 tok/s
- OrcaOrcaCyber: MiMo V2.6 Flash Abliterated / Uncensored2026-10-26$0.80 / $1.60 per 1M tokens · 43 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 · 192 tok/s
- minimaxMiniMax: M3.1 Flash2026-10-1929Intelligence
- OrcaOrca: OrcaCyber Ultra 1.02026-10-17$5.00 / $10.00 per 1M tokens · 130 tok/s
- deepseekDeepSeek: DeepSeek V4.1 Flash 10172026-10-1733Intelligence
- DeepSeekDeepSeek: DeepSeek V4.1 Flash2026-10-1740Intelligence
- OrcaOrca: OrcaCyber Zero 1.52026-10-10$3.00 / $7.50 per 1M tokens · 135 tok/s
- openaiOpenAI: GPT-6.1 Sol2026-09-2955Intelligence
- anthropicAnthropic: Claude Sonnet 5.52026-09-2855Intelligence
- typesafeTypeSafe: Jev 1.132026-09-24$0.04 / $0.00 per 1M tokens · 182 tok/s
The first thing that separates MAI Decision 1 — the model Microsoft shipped at Ignite as Microsoft-Decision-1.1 — from Nimble v3, filed on Hugging Face as Bespoke-Nimble-9B-v3, is not a benchmark. It is a licence clause. Nimble v3 is a LoRA adapter released under CC BY-NC 4.0, which the model card glosses as "may not be used for commercial purposes". Microsoft's model is a hosted endpoint you rent per token, so the licence question never arises and the self-hosting question never arises either. One of these you can put in a product; the other you can put in a notebook. Everything else in this comparison follows from that fork, including why their published scores should not be subtracted from each other.
What each one actually is, mechanically
Microsoft-Decision-1.1 is a Foundry deployment, live since 17 November 2026, post-trained by Microsoft on Gemma 4 26B-A4B. You send a state and typed questions to a decision endpoint, you get calibrated probabilities back, and you are billed $0.042 per million input tokens on Global Standard or $0.05 in the EU and APAC data zones. There is no download, no fine-tuning path, and no self-hosting option in the model card. Latency is 123 ms per task on Artificial Analysis's measurement; context is 32,768 input tokens; it reads images as well as text.
Nimble v3 is a five-file PEFT adapter for Qwen3.5-9B, uploaded on 30 September 2026 with no accompanying announcement, no blog post and no release note from Bespoke Labs. The repository holds a README, an adapter config, adapter weights, a schema config and a gitattributes file — that is the complete list. You bring the base model, the inference engine and the GPU yourself. The companion dataset repository carries the full Decision Index run, the engine and the reproduction instructions, which is more transparency than most vendors offer and also more work than most buyers expect.
• Weights — hosted only, no download vs a 20 GB BF16 adapter you assemble onto Qwen3.5-9B
• Licence — Microsoft service terms vs CC BY-NC 4.0, non-commercial, with an Apache 2.0 base that does not loosen the adapter's restriction
• Interface — a decision endpoint with typed questions and probabilities vs your own inference code, pinned to prompt files and option-code tokens by hash
• Context — 32,768 tokens vs 32,768 tokens, with oversized requests refused rather than truncated
• Images — up to 8 per question, 280 to 1,568 pixels per side vs text only
• Price — $0.042 per million tokens, input only vs hardware, electricity and an engineer
• Published score — 67.78 on AA Decision Index v2.1 vs 56.88 on Decision Index 0.2.1

The two numbers are not on the same ruler
That last bullet is the one that gets abused in comparison posts, so it is worth being blunt: 56.88 and 67.78 are not comparable. Nimble v3's figure comes from Decision Index 0.2.1, the open kit Bespoke Labs runs itself, on its own harness, with its own reference implementation. Microsoft's 67.78 comes from Artificial Analysis Decision Index v2.1, a different composition of evaluations with different weights, a different scoring function and a different set of models in the field.
Both boards measure the same family of behaviour — you pick among bounded options and you are scored on accuracy discounted by how badly your stated confidence missed. But a six-evaluation weighted composite with a symmetry term is not the same instrument as a public kit run, and subtracting one from the other produces a 10.9-point "gap" that is mostly a difference in rulers. Anyone showing you that subtraction is either careless or selling something.
What can be said without overreach is directional. Nimble v3's own card reports one number and no breakdown, in a family whose earlier checkpoint published a much fuller table: 90.12% agreement with reference labels against 93.21% for Jev 1.13.0 on 324 held-out contrastive pairs, and 75.7% against 76.0% on 3,880 human-labelled decisions across thirteen public datasets. Those are the first Nimble's numbers, not v3's, and the project labels them as its own measurements. Decision 1.1, by contrast, carries an independent 67.78 with six published sub-scores and a Brier of 0.160 on a board that also ranks its competitors the same way. If you need a number you can defend in a procurement review, only one of these gives you one.

The licence is the whole decision, and it is worse than it looks
CC BY-NC 4.0 on the adapter means no commercial use. That is not a grey area you can reason around, and the obvious workaround does not work: the base model, Qwen3.5-9B, is Apache 2.0, and Bespoke's card says so explicitly — but a permissive base does not loosen a non-commercial adapter, and the adapter is where the decision behaviour lives. Strip the adapter and you have a general 9B model that is not a decision model at all.
There is a second, less discussed problem. The GitHub repository for the Nimble project reports no licence at all through the API, which means the training and curation code has no declared terms either. If your legal team asks what you may do with the pipeline that produced the weights, the answer is that the project has not said.
Set that against Microsoft's model: you cannot download it, so you cannot fork it, but you also cannot accidentally build a product on a licence that forbids the product. For a team deciding between the two, the practical framing is that one is a purchase and the other is a research artefact. If the workload is commercial, the comparison ends here and the hosted model wins by default. If it is not — an internal evaluation harness, an academic study, a prototype you intend to replace — then the adapter is genuinely interesting and its 20 GB footprint means it runs on a single GPU you may already own.
Deployment reality on each side
Running Nimble v3 correctly is not a matter of loading the adapter and chatting to it. The dataset card is explicit that the engine builds the prompts its schema config pins by hash and reads each request with the attention pattern the model was trained with; loading the weights as a plain chat model, or generating text, does not reproduce its answers. The engine refuses to load if the prompt files or option-code tokens differ from training. You will need CUDA, BF16, torch 2.8, transformers 5.17.0, peft and two attention kernels, and requests above the context limit are rejected rather than shortened.
That is a real engineering commitment and it buys you something the hosted model cannot: the ability to run with no network egress, on your own hardware, with the option of inspecting every prompt. For regulated environments where a model call cannot leave the building, it is the only one of the two that is even a candidate — and the non-commercial licence may still rule it out, which is the trap.
On the Microsoft side the cost is different and smaller. A Foundry deployment, an endpoint, a deployment name, a key or an Entra token. Co-locate the Foundry resource with your application if latency matters; use a data-zone deployment if inference has to stay in-region, at $0.05 per million tokens rather than $0.042. Your integration is an HTTPS call with a JSON body, and the response shape is probabilities you threshold.
• Time to first call — under an hour on Foundry vs a GPU setup, a pinned toolchain and a reproduction run
• Failure mode — a rate limit or a region outage vs a CUDA OOM or a kernel version mismatch
• Upgrade path — none, the vendor ships a new version vs a new adapter tag you can pin or ignore
• Exit path — none, hosted only vs the weights stay where they are, subject to the licence
Where the successor complicates the story
Nimble v3 is not Bespoke Labs' current model, and that matters for anyone evaluating it now. Nimble-v4-max, uploaded on 3 November 2026, is a full fine-tune of Qwen3.8-27B rather than an adapter, published with a tag history and a benchmark table: 72.74 on Decision Index 0.3 public, up from 68.46 at v4.0, across 43 benchmarks whose per-benchmark scores ship in a JSON file. It needs about 55 GB in BF16 and one 80 GB GPU. It is also CC BY-NC 4.0.
So the honest read on Nimble v3 today is that it is the previous generation of a line that has already moved twice, and its one published score is on a kit version the current model has left behind. Comparing a vendor's three-week-old hosted flagship against a quietly-uploaded adapter two revisions old is not a fair fight on either axis. The product question is genuinely Microsoft-Decision-1.1 against Nimble-v4-max, and even there the licence answer is unchanged.
One key, when you are still deciding
Neither model is on OrcaRouter's catalogue, and this piece is not going to suggest otherwise: Microsoft-Decision-1.1 is Foundry-only, and a CC BY-NC adapter is not something a commercial router can serve. What the catalogue does carry is the neighbourhood, which is where most evaluations actually start.
The neighbourhood is still useful, because both sides of this matchup define themselves against the same rival. Jev 1.14 is that rival: Microsoft-Decision-1.1 beats it by 1.55 points on the independent board, and Bespoke's earlier checkpoint measured itself against Jev 1.13.0 in its own README. Jev 1.14 answers at $0.042 per million tokens in and out. The base Gemma 4 26B A4B, which is what Microsoft post-trained onto, is routable at $0.25 per million input and $2.90 output. Put Jev and a general model on your own labelled traffic behind one endpoint, with automatic failover, and you learn in an afternoon whether a decision model earns its place in the path — before a procurement cycle, and before a GPU purchase. If it does, the router stays in front of whatever decision model you deploy, so one provider's rate limit is a reroute rather than an incident.

What that harness will not tell you is how Decision 1.1 itself behaves on your data. For that you need a Foundry deployment and your own labels, and there is no shortcut around the labelling.
How to choose
Take Microsoft-Decision-1.1 if the workload is commercial, if you want a calibrated score you can point at an independent board, if you need image input, or if you would rather rent 123 ms than own a GPU pipeline. That is most teams.
Consider Nimble v3 only if all three of these hold: the use is non-commercial, the model call cannot leave your infrastructure, and you are prepared to run the project's own engine rather than a generic loader. If any one fails, the licence or the engineering has already made the decision for you — and if you are choosing a Nimble at all today, look at Nimble-v4-max first, because v3's single published score is the oldest number in the comparison.
The thing to resist is the tidy table. A 56.88 against a 67.78 looks like a verdict and is not one; it is two different kits, two different ages and two different licences dressed as a horse race. The model that ships in your product is the hosted one, and the model that runs in your lab is the open one — that is the real split, and no benchmark moves it.
