Technical Specifications
Table with columns: Parameter, Specification| Parameter | Specification |
|---|
| Total Base Parameters | 8.9B (Vision tower preserved; Multi-Token Prediction excised) |
| Foundation Architecture | Hybrid Recurrent-Attention Causal LM (Qwen3_5ForCausalLM) |
| Recurrent Topology | Gated DeltaNet linear-attention layers interleaved with standard attention |
| Attention Configuration | Grouped-Query Attention (GQA) with YaRN RoPE scaling |
| Context Window | 262,144 tokens native |
| Native Precision | bfloat16 |
| Primary Capabilities | Agentic coding and software engineering |
Direct linear blending, spherical interpolation (SLERP), and uniform DARE perturbations degrade hybrid DeltaNet networks. Linear attention recurrence relies on strictly contractive transition states; arbitrary parameter shifts introduce runaway eigenvalues that destroy long-range memory and break special control tokens.
To prevent degradation, the parameter integration is governed by non-interfering geometric constraints.
1. Delta Task Disentanglement
Let W0 denote the underlying base model parameters (Qwen3.5-9B), WMiMo the primary anchor, and WHERO the donor. The task vectors are defined as:
ΔMiMo=WMiMo−W0
ΔHERO=WHERO−W0
2. Spectrally Tapered Bi-Subspace Guard
To guarantee that HERO does not overwrite the fundamental representations discovered during MiMo's distillation, HERO's update is projected orthogonally away from the dominant singular directions of MiMo's task vector.
For any weight matrix with singular value decomposition:
ΔMiMo≈UkΣkVkT
The overlapping projection tensor is computed across rank k = 32 and tapered according to spectral energy:
Ω=(τdiag(Σk))2
ΔHERO⊥=ΔHERO−Uk(Ω⊙(UkTΔHEROVk))VkT
Components of the agentic donor that directly compete with MiMo's primary singular subspace are eliminated, allowing new procedural capabilities to occupy unutilized parameter dimensions.
3. Sign Consensus Damping
When fine-tuning objectives pull individual weights in opposite directions, blind averaging causes catastrophic cancellation and syntax corruption. For coordinates where updates conflict:
sign(ΔMiMo(i,j))=sign(ΔHERO⊥(i,j))
An antagonistic penalty coefficient Φ is evaluated against relative weight salience:
Φ(i,j)=1.0−0.50⋅(∣ΔMiMo(i,j)∣+∣ΔHERO⊥(i,j)∣+ϵ∣ΔMiMo(i,j)∣)
ΔHERO(i,j)=Φ(i,j)⋅ΔHERO⊥(i,j)
Antagonistic donor noise is suppressed, ensuring that settled code generation rules are never inverted.
4. DARE Energy-Preserved Trimming
To eliminate low-magnitude parameter drift, the bottom 4% energy of the sanitized update is zeroed out. The surviving parameters are rescaled by the inverse square root of the retained energy ratio to preserve exact variance:
ΔHERO=Mask(ΔHERO,p=0.96)⋅0.961.0
5. Bell-Plateau Depth Routing
Layer-wise integration follows a smooth transition profile across the network's depth index l over total layers L:
g(l)=⎩⎨⎧0.50+0.50⋅S(0.20Ll)1.001.00−0.30⋅S(0.20Ll−0.80L)0≤l<0.20L0.20L≤l≤0.80L0.80L<l≤L
where S(t) represents the smooth cubic step function:
S(t)=3t2−2t3
- Early Layers (0% to 20%): Lower graft gain (0.50) safeguards token embeddings and initial sensory projections.
- Mid Layers (20% to 80%): Full-strength injection (1.00) transforms core reasoning and planning pathways.
- Late Layers (80% to 100%): Scaled down (0.70) to protect tool-call formatting, stopping tokens, and output logits.
6. Dual Trust Region Clamping
To prevent distribution collapse, the effective step size is bound simultaneously by Frobenius norm ratio ρ and spectral radius σmax:
s=min(smax,ρ⋅∥ΔHERO∥F∥ΔMiMo∥F,κ⋅σmax(ΔHERO)σmax(ΔMiMo))
Final parameter state:
Wfinal=WMiMo+(s⋅g(l))⋅ΔHERO
Subsystem Protection Policies
Table with columns: Subsystem, Target Tensors, Policy, Objective| Subsystem | Target Tensors | Policy | Objective |
|---|
| Embeddings & Heads | embed_tokens, lm_head | Strict Anchor Lock | Prevents vocabulary misalignment and maintains exact output logit distribution. |
| Vision Subsystem | visual.* | Preserved Intact | Keeps multimodal parsing and visual document understanding fully functional. |
| Recurrent Core | linear_attn.* | Conservative Gain (0.80) | Protects Gated DeltaNet state-space stability across 128k context steps. |
| Attention Routing | self_attn.q_proj, k_proj | Attenuated Gain (0.85) | Prevents attention-head defocusing during long-sequence tool tracking. |
| Value & Projections | self_attn.v_proj, o_proj | Elevated Gain (0.95) | Facilitates smooth transfer of semantic agentic representations. |
| MLP Blocks | mlp.down_proj, up_proj | Full Strength (1.00) | Maximizes transfer of practical code implementation heuristics. |
| Outlier Coordinates | |W| > 8 · rms(W) | Coordinate Shield | Protects critical emergent outlier features from perturbation. |
Chat Template & Agentic Protocol
This model integrates the Improved Chat Template for Qwen 3.x directly into its tokenizer configuration and Jinja runtime environment.
Protocol Highlights
- Multi-Tier Chain-of-Thought Control: Isolates internal reasoning paths within explicit
<thought>...</thought> delimiters prior to tool calling or terminal emission.
- Dual Tool Schemas: Fully supports both JSON-RPC function payloads and structured XML tool execution tags, enabling zero-shot integration with platforms like Browser-Use, OpenHands, and native agent loops.
- Anti-Repetition Circuit Breaker: Suppresses cyclical generation traps when linters, compilers, or shell environments return non-zero exit codes.
- Strict Whitespace & Prompt Hygiene: Eliminates redundant framing tokens, conserving the active context budget for long-sequence programming and multi-turn debugging.
Recommended Generation Parameters
For repository-level software engineering, bug localization, and automated tool use:
Table with columns: Parameter, Recommended Value, Description| Parameter | Recommended Value | Description |
|---|
| Temperature | 0.6 | Balances strict syntactic validity with algorithmic path exploration. |
| Top-P | 0.95 | Eliminates low-probability token tails while preserving alternative logic paths. |
| Top-K | 20 | Restricts token candidate pools to prevent architectural syntax drift. |
| Min-P | 0.0 (Off) | Disabled in favor of explicit Top-K / Top-P governance. |
| Repetition Penalty | 1.0 (Off) | Disabled to prevent syntax degradation in repetitive code patterns (indentation, braces). |
| Presence Penalty | 0.0 | Prevents naming mutations across long-context symbol resolution. |
Inference and Deployment
Serving via vLLM
Thanks to the excision of multi-token prediction layers, the model deploys natively without custom engine flags:
vllm serve pragmaticcs/MiMo-HERO-UIUX-Designer-9B \ --dtype bfloat16 \ --max-model-len 32768 \ --gpu-memory-utilization 0.90 \ --enable-auto-tool-choice \ --tool-call-parser hermes \ --enable-reasoning \ --reasoning-parser qwen3
Serving via SGLang
python -m sglang.launch_server \ --model-path pragmaticcs/MiMo-HERO-UIUX-Designer-9B \ --port 30000 \ --host 0.0.0.0 \ --context-length 32768
Python API Execution
import torchfrom transformers import AutoModelForCausalLM, AutoTokenizer model_id = "pragmaticcs/MiMo-HERO-UIUX-Designer-9B" tokenizer = AutoTokenizer.from_pretrained(model_id)model = AutoModelForCausalLM.from_pretrained( model_id, torch_dtype=torch.bfloat16, device_map="auto") conversation = [ { "role": "user", "content": "Create a beautiful one-file landing page for a travel agency using HTML, CSS and JavaScript.", },] inputs = tokenizer.apply_chat_template( conversation, add_generation_prompt=True, enable_thinking=True, return_tensors="pt").to(model.device) outputs = model.generate( inputs, max_new_tokens=65536, temperature=0.6, top_p=0.95, top_k=20, do_sample=True) response = tokenizer.decode(outputs[0][inputs.shape[-1] :], skip_special_tokens=True)print(response)
Citation & References
@article{yu2024dare, title={Language Models are Super Mario: Absorbing Abilities from Homologous Models as a Free Lunch}, author={Yu, Le and Yu, Bowman and Yu, Haiyang and Huang, Fei and Li, Yongbin}, journal={arXiv preprint arXiv:2311.03099}, year={2023}} @inproceedings{yadav2023ties, title={Resolving Interference When Merging Models}, author={Yadav, Prateek and Tam, Derek and Choshen, Leshem and Raffel, Colin and Bansal, Mohit}, booktitle={Advances in Neural Information Processing Systems (NeurIPS)}, volume={36}, pages={7093--7115}, year={2023}} @article{yang2024deltanet, title={Gated Delta Networks: Improving Mamba2 with Delta Rule for Linear Recurrent Attention}, author={Yang, Songlin and Wang, Bailin and Shen, Yikang and Panda, Rameswar and Kim, Yoon}, journal={arXiv preprint arXiv:2412.06464}, year={2024}}