MiMo-V2.6-Distill-Qwen-9B review, pricing and limits
A small, self-hostable research checkpoint in Xiaomi's MiMo-V2.6 family: an MIT-licensed SFT distillation of Qwen3.5-9B for agentic coding, cybersecurity, and visual tasks.
- Best for
- Researchers experimenting with agentic RL on a small open base
- Not for
- Production workloads needing a hosted, supported API
Should you use MiMo-V2.6-Distill-Qwen-9B?
This 9B checkpoint trains on 77.4 billion tokens split across code, cybersecurity, general agent tasks, and visual coding, with 20 community finetunes already published. It suits researchers and self-hosters wanting a small, MIT-licensed agentic base, not teams needing a hosted API or flagship-level intelligence.
Best for
- Researchers experimenting with agentic RL on a small open base Xiaomi releases it explicitly as a starting point for open agentic-RL research, and it already has 20 community finetunes on Hugging Face.
- Self-hosting a small MIT-licensed coding/agent model 9B dense parameters in BF16 safetensors, small enough for a single modern GPU, unlike the 1.02T-parameter Pro-RL.
- Teams already invested in the Qwen3.5 ecosystem Built as a direct SFT fine-tune of Qwen3.5-9B, so it inherits that model's tokenizer and chat template conventions.
Not for
- Production workloads needing a hosted, supported API Xiaomi's pay-as-you-go API pages list only MiMo-V2.6-Pro and MiMo-V2.6-Flash; this checkpoint is self-host only.
- Tasks needing capability close to the RL-trained flagships It is a supervised fine-tune, not an RL-trained checkpoint, and Xiaomi frames it as a research starting point rather than a Pro-RL or Flash-RL competitor.
MiMo-V2.6-Distill-Qwen-9B pricing
- Free tier
- No
No hosted API exists for this checkpoint as of this listing; Xiaomi's pay-as-you-go pricing pages list only MiMo-V2.6-Pro and MiMo-V2.6-Flash. Running MiMo-V2.6-Distill-Qwen-9B means self-hosting the open weights, so the real cost is whatever compute you provide, not a per-token vendor rate.
Self-hosting requirements
- Self-hosted
- Yes
- Recommended GPU
- Xiaomi does not publish an official minimum. As a rough, non-vendor estimate: 9B parameters in BF16 need about 18GB of VRAM for the weights alone, before KV cache, so a single modern 24GB+ GPU is plausible for short-context use.
What MiMo-V2.6-Distill-Qwen-9B can do
MiMo-V2.6-Distill-Qwen-9B is a 9-billion-parameter dense model Xiaomi released by supervised fine-tuning Qwen3.5-9B on MiMo-generated agentic data. Released September 21, 2026 under the MIT license, it raises SWE-bench Verified from its base model's 60.0 to 61.1 and ships purely as open weights, with no hosted API.
What goes in and what comes out
Reads
- text
- image
Writes
- text
Capabilities
- Function callingSupported
- Tool useSupported
- VisionSupported
- Audio inputNot supported
- Audio outputNot supported
- Code executionNot supported
- Structured outputNot supported
- Video inputNot supported
- Web browsingNot supported
- Reasoning modes
- standard, extended-thinking
Key features
9B dense SFT of Qwen3.5-9B
Supervised fine-tune of Alibaba's Qwen3.5-9B on 77.4 billion tokens of MiMo-generated agentic data, released as an open research starting point rather than a flagship.
Covers four agentic domains
The SFT mixture splits across code (29.9%), general agent tasks (28.5%), visual coding (27.4%), and cybersecurity (14.2%).
Thinking on/off toggle
Supports the same enable_thinking chat-template switch as the rest of the MiMo-V2.6 family, separating reasoning_content from the final answer.
Vision-capable, text output
Handles image-text-to-text input, inherited from the Qwen3.5-9B base, but Xiaomi has not documented video or audio input for this checkpoint the way it has for Pro-RL and Flash-RL.
Active open-source community
A downstream community of finetunes, merges, and quantizations has already formed around the checkpoint on Hugging Face, unusually fast for a research-framed release.
Known limitations
- A supervised fine-tune rather than an RL-trained checkpoint, so it trails the family's own flagships by design, not just by size.
Who it is for
- Researchers building on agentic reinforcement learning who want a small open starting checkpoint
- Developers already using Qwen3.5-9B who want MiMo-style agentic behavior without switching base models
- Hobbyists and small teams who want to self-host a capable coding/agent model on a single GPU
- Teams that do not need a hosted API or flagship-level intelligence
Supported languages
- English
- Chinese
MiMo-V2.6-Distill-Qwen-9B benchmarks
Standout result
SWE-bench Verified (avg@3), up from the Qwen3.5-9B base model's 60.0
Source: Xiaomi MiMo-V2.6-Distill-Qwen-9B Hugging Face model card, verified 2026-09-28Where it sits
In the bottom third on SWE-bench Verified (rank 30 of 39), and one of 89 whose vendor states it does not train on customer data.
| Measure | This model | Rank | Median |
|---|---|---|---|
| SWE-bench Verified% solved | 61.1% | #30 / 39 | peer median 76.5% |
Ranks are against GA models on Hokai that publish the same figure; ties share a rank.
A benchmark is an exam, not the job. Scores transfer unevenly between tasks, so weigh the one closest to your workload and read every figure with its source.
Strengths and limits
Strengths
- MIT license with no commercial-use restrictions, plus an active downstream community already building finetunes and quantizations on Hugging Face.
- Improves meaningfully over its base model on every benchmark Xiaomi reports, for example 37.1 vs 27.0 on Terminal-Bench 2.1.
- Improves over its Qwen3.5-9B base on every benchmark Xiaomi reports, for example Terminal-Bench 2.1 rising from 27.0 to 37.1.
- At 9 billion dense parameters it needs a fraction of the hardware the 1.02T-parameter Pro-RL calls for.
- Already has 20 community finetunes, 7 merges, and 54 quantizations on Hugging Face, well ahead of Pro-RL's 11 quantizations.
Limits
- A supervised fine-tune rather than a checkpoint trained with reinforcement learning, so its ceiling is well below the family's two flagships.
- No published context window for this specific checkpoint, unlike the two hosted flagships.
MiMo-V2.6-Distill-Qwen-9B alternatives
The nearest AI models on Hokai for the same job.
| Model | Lab | Context | In / out per 1M | Reads and writes |
|---|---|---|---|---|
| MiMo-V2.6-Distill-Qwen-9B This page | Xiaomi Corporation | -- | -- | text, image → text |
 Claude Fable 5Superseded by Fable 5.1 and Fable 5.2 but still sold on the Claude API: Anthropic's first generally available Mythos-class model, released June 2026, with a 1M-token context. | Anthropic | 1M | $10 / $50 | text, image, pdf → text |
| Claude Fable 5.1Superseded by Claude Fable 5.2 in November 2026 but still sold on the Claude API: Anthropic's second Mythos-class model, released Sept 1, 2026, for long-horizon coding and research. | Anthropic | 1M | $10 / $50 | text, image, pdf → text |
 Gemini Decision 1Google DeepMind's decision model scores many questions over one multimodal input of up to 262,144 tokens and bills zero output tokens. | 262K | $0.10 / $0 | text, image, audio, video, pdf → text |
MiMo-V2.6-Distill-Qwen-9B specs
- Architecture
- Dense Transformer (Qwen3.5 base, SFT fine-tuned)
- Mixture of experts
- No
- Parameters
- 9B (dense)
Training data
A 77.4-billion-token supervised fine-tuning mixture (27.2B loss-bearing) generated by Xiaomi's own larger MiMo models: code 23.2B tokens (29.9%), cybersecurity 11.0B (14.2%), general agent tasks 22.0B (28.5%), visual coding 21.2B (27.4%). This is distillation data, not raw pretraining text; the underlying Qwen3.5-9B base's own pretraining data is documented separately by Alibaba.
Weight formats
- Safetensors (BF16)
SDKs
- Python
Prompting patterns
Show 2 prompting patterns
Explicit thinking toggle
extra_body={"chat_template_kwargs": {"enable_thinking": true}}The model card's own quickstart enables extended reasoning per request; the response separates reasoning_content from the final answer.
Use the checkpoint's own bundled chat template
Serve with the tokenizer shipped in the MiMo-V2.6-Distill-Qwen-9B repo rather than a generic Qwen3.5 template
The model card notes the checkpoint includes its own tokenizer and MiMo v2.6 chat template; a mismatched template risks broken reasoning-tag parsing.
Known quirks
Show 3 known quirks
Trained on synthetic MiMo data, not raw web text
The 77.4B-token SFT mixture is data Xiaomi's own larger MiMo models generated, not independently documented pretraining data; its behavior reflects whatever those teacher models produced.
Workaround: Evaluate the model against your own held-out tasks before trusting it outside the four domains the SFT mixture covers: coding, cybersecurity, general agent work, and visual coding.
No published context window for this checkpoint
Unlike Pro-RL and Flash-RL, which both publish a 1M-token context window, this self-host-only checkpoint's model card does not state a context length.
Workaround: Check Qwen3.5-9B's own published context window as a practical starting point, and verify empirically with your own long-context tests before relying on it.
Needs a specific SGLang build and the checkpoint's own chat template
The model card's quickstart requires a recent SGLang build with Qwen3.5 support plus the checkpoint's bundled tokenizer and chat template, not a generic Qwen or MiMo template.
Workaround: Follow the model card's exact SGLang invocation with --reasoning-parser mimo rather than reusing a serving config built for a different Qwen or MiMo checkpoint.
MiMo-V2.6-Distill-Qwen-9B facts
- Deployment
- Self-hosted
- Trains on your data
- No
- Licence
- MIT · Open source
- Safety posture
- Balanced
- Released
- 21 Sep 2026
- Current version
- MiMo-V2.6-Distill-Qwen-9B
- Family
- MiMo V2.6
- Variant
- Multimodal
- Data retention
- There is no hosted API for this checkpoint, so no user data is sent to Xiaomi; it is downloaded and run entirely on the operator's own infrastructure.
Safety controls
- Rlhf
- No
- System prompt
- Yes
- Constitutional AI
- No
- Content filter API
- No
- Moderation endpoint
- No
- Red teaming
- Not independently documented for this checkpoint. It is a supervised fine-tune on Xiaomi-generated data rather than an RL-trained model, so the alignment methods described for Pro-RL and Flash-RL do not directly apply.
Data governance
- Trains on inputs
- No
Version history
MiMo-V2.6-Distill-Qwen-9B · 21 Sep 2026
Initial release: SFT of Qwen3.5-9B on 77.4B tokens of MiMo-generated data. Improves SWE-bench Verified from the Qwen3.5-9B base's 60.0 to 61.1.
What changed
- Release MiMo-V2.6-Distill-Qwen-9B released as an open agentic-RL research base source
Questions about MiMo-V2.6-Distill-Qwen-9B
Is MiMo-V2.6-Distill-Qwen-9B free to use?
Yes. It ships under the MIT license, which permits commercial use, and there is no hosted API or per-token charge: you download the open weights from Hugging Face and run them on your own infrastructure. The only cost is whatever compute you provide.
What separates MiMo-V2.6-Distill-Qwen-9B from its Qwen3.5-9B base model?
Fine-tuning on MiMo-generated agentic data lifts SWE-bench Verified from 60.0 to 61.1, SWE-bench Pro from 32.0 to 44.6, and Terminal-Bench 2.1 from 27.0 to 37.1, per Xiaomi's own reported figures. Qwen3.5-9B itself remains available separately from Alibaba as an untuned base.
Is MiMo-V2.6-Distill-Qwen-9B open source or proprietary?
It is open source under the MIT license, and the weights are published on Hugging Face. At 9 billion dense parameters, hardware requirements are far below the 1.02-trillion-parameter Pro-RL checkpoint in the same family, which needs an 8-16 GPU cluster to serve.
What happens to your data when you use MiMo-V2.6-Distill-Qwen-9B?
Nothing leaves your own infrastructure, because there is no hosted API: you download the weights and run inference yourself, so Xiaomi never sees prompts or outputs. That is a different posture from the hosted Pro-RL and Flash-RL checkpoints, which do have a live API.
What is MiMo-V2.6-Distill-Qwen-9B not good at?
It is a supervised fine-tune, not an RL-trained checkpoint, so it should not be expected to approach Pro-RL or Flash-RL on agentic benchmarks or the Artificial Analysis Intelligence Index. It also has no documented context window, no hosted API, and no confirmed video or audio input, unlike its two larger siblings.
About MiMo-V2.6-Distill-Qwen-9B
MiMo-V2.6-Distill-Qwen-9B is a 9-billion-parameter dense model Xiaomi's MiMo team released alongside the flagship MiMo-V2.6-Pro-RL and MiMo-V2.6-Flash-RL checkpoints on September 21, 2026, under the MIT license. Unlike its two RL-trained siblings, it is a supervised fine-tune of Alibaba's Qwen3.5-9B base model on MiMo-generated data, not a reinforcement-learning checkpoint in its own right. Xiaomi frames it as a starting point for open research into agentic reinforcement learning, covering coding, general-purpose agent tasks, visual coding, and cybersecurity.
Read the full background
Xiaomi's own evaluation table compares the SFT checkpoint against its Qwen3.5-9B base rather than competing 9B models. On SWE-bench Verified it improves from 60.0 to 61.1 (avg@3), and on SWE-bench Pro from 32.0 to 44.6. On Terminal-Bench 2.1 it rises from 27.0 to 37.1 (avg@1). Larger jumps appear on Xiaomi's own internal evaluation sets, which are not independently verifiable: an internal coding set from 19.5 to 51.6, and an internal general-agent set from 28.5 to 62.2.
As a Qwen3.5-9B derivative, the model inherits that architecture's context handling and tokenizer rather than Pro-RL's own sparse-MoE design; Xiaomi's model card does not publish a specific context window for this checkpoint the way it does for the two hosted flagships. It ships as BF16 safetensors and supports the same explicit thinking on/off toggle used across the MiMo-V2.6 family.
Input is text and image, listed under Hugging Face's image-text-to-text task and inherited from the Qwen3.5-9B base's own vision support. There is no evidence of video or audio input: unlike Pro-RL and Flash-RL, its Hugging Face tags do not list video-understanding or audio capability. Output is text only.
The supervised fine-tuning mixture totals 77.4 billion tokens, of which 27.2 billion are loss-bearing. By domain: code is 23.2 billion tokens (29.9%), cybersecurity 11.0 billion (14.2%), general agent tasks 22.0 billion (28.5%), and visual coding 21.2 billion (27.4%). This is distillation data Xiaomi's own larger MiMo models generated, not raw pretraining text.
There is no hosted API for this checkpoint; Xiaomi's pay-as-you-go product pages list only MiMo-V2.6-Pro and MiMo-V2.6-Flash. It is distributed purely as open weights on Hugging Face, with a documented SGLang quickstart, and broad downstream community activity: 20 finetunes, 7 merges, and 54 quantizations were already published by others at listing time, alongside 9,994 downloads in the preceding month.
As with the two flagship checkpoints, Xiaomi has not published a MiMo-V2.6-specific training-data summary or system card covering this distilled model. Because it trains on Xiaomi's own MiMo-generated data rather than raw web text, its behavior is shaped by whatever the teacher models produced, which is not independently documented.
MiMo-V2.6-Distill-Qwen-9B suits researchers and hobbyists who want a small, self-hostable, MIT-licensed agentic model to experiment with or fine-tune further, and teams building on Qwen3.5's ecosystem who want MiMo-style agentic behavior without the trillion-parameter Pro-RL checkpoint. It is a weaker choice for anyone needing a supported hosted API, a long context window, or capability close to the much larger RL-trained siblings.
The checkpoint sits inside Xiaomi's 5-item MiMo-V2.6 Hugging Face collection alongside Pro-RL and Flash-RL, and Xiaomi frames it as a base for others to run their own agentic reinforcement learning on top of, rather than a finished product. As with the rest of the family, Xiaomi has already previewed MiMo-V3 as the next generation in the line.
Sources and guides
Official sources
- MiMo-V2.6-Distill-Qwen-9B model page Vendor
- Licence text Licence
Specifications on this page (context window, token prices, benchmark scores, licence) are read from Xiaomi Corporation's own model card and documentation. The verdict is Hokai's reading of them, not the provider's claim.
Lists that include MiMo-V2.6-Distill-Qwen-9B
The provider
More AI Models on Hokai
Shieldstral 1.0Mistral's open-weight multimodal safety classifier, released 4 August 2026, that screens text and images against policies you write.
- Claude Mythos 5.2Anthropic's November 2026 Mythos model: Fable 5.2's weights with relaxed cyber and biology safeguards for verified teams, with a 1M-token context.
ERNIE-5.5-Flash-PreviewBaidu's text-only preview mixture-of-experts model, served through Qianfan on OpenRouter with a 1M-token context window and tool calling.
Gemini 4 CarbonGoogle DeepMind's Gemini 4 Carbon, open to paying Gemini API developers since 10 November: a reasoning model for agentic coding, research and cyber defense.
JevTypeSafe's first System One Model, launched in 2026: skips text generation for typed, calibrated decisions software can act on directly.Nex-N2.6Nex AGI's hosted agentic model for computer use and coding, with image input, tool calling and about 321B parameters, served first-party on OpenRouter.