Qwen3.6-35B-A3B — Claude 4.7 Opus Reasoning Distilled — PrismaQuant 4.75 bpp
Mixed-precision quantization of
Qwen3.6-35B-A3B-Claude-4.7-Opus-Reasoning-Distilled produced by
PrismaQuant — a per-Linear sensitivity-driven allocator that chooses each Linear module's format individually under a total-bit budget.
The source model is a reasoning distillation of Qwen3.6-35B-A3B trained on Claude 4.7 Opus chain-of-thought traces. It retains the full multimodal (vision + text) and MTP speculative-decoding head of the base architecture, with added reasoning capability. Enable thinking mode at inference time (see Serving below).
Why "every layer refracts into a different format": a naive uniform NVFP4 either leaves disk on the table (keeping everything BF16 "to be safe") or loses quality (quantizing sensitive layers to 4-bit). PrismaQuant measures the actual Fisher-weighted MSE for every (Linear, format) pair and runs a multi-choice knapsack under a total-bit budget, so every bit lives where it buys the most likelihood.
At a glance
| Metric | BF16 source | This artifact |
|---|
| Size on disk | ~70 GB | ~22 GB |
| Average bits per param | 16 | 4.75 |
| Reasoning (Claude 4.7 Opus distillation) | ✓ | ✓ |
| Multimodal (vision + text) | ✓ | ✓ |
| MTP speculative decoding heads | ✓ | ✓ |
Loads in vLLM (stock compressed-tensors) | ✓ | ✓ |
| Runtime backend | any | vLLM only |
Precision mix
This checkpoint uses three precisions, selected per-Linear by the allocator from measured sensitivity — not chosen uniformly. The visual encoder is kept entirely at BF16 (--visual-format=BF16), preserving full visual fidelity at the cost of a small disk overhead vs a quantized visual tower.
Allocator result (text body, 401 Linears):
| Format | Bits | Count (body alloc) |
|---|
| NVFP4 | 4.50 | 121 Linears |
| MXFP8_E4M3 | 8.25 | 48 Linears |
| BF16 | 16.00 | 109 Linears |
Export recipe (all modules, 511 entries including MTP + visual passthrough):
| Format | W | A | Use | Count |
|---|
| NVFP4 | 4-bit (FP4, group_size=16 with per-group FP8 scale + per-tensor global) | 4-bit (dynamic) | Bulk MoE per-expert + medium-sensitivity dense Linears | 154 entries (71 dense + 80 per-expert + MTP) |
| MXFP8 | 8-bit (E4M3, group_size=32 with per-group E8M0 scale) | 8-bit (dynamic) | High-sensitivity dense Linears the allocator won't risk at 4-bit | 76 Linears |
| BF16 | 16-bit | 16-bit | Router, norms, biases, embed / lm_head, pos_embed, full visual encoder (110 Linears) | 281 entries |
The allocator couples gate_up_proj / down_proj siblings so they share one per-tensor global scale (vLLM's FusedMoE requires this), and fused attention siblings (q_proj/k_proj/v_proj) share one per-tensor global scale so the packed qkv_proj loads without the "accuracy mismatch" warning. 112 fused-sibling groups were resolved during allocation.
61 Linears were ineligible for MXFP8_E4M3 due to kernel shape constraints and were restricted to the NVFP4 / BF16 choice.
Activation-aware passes applied during export
On every NVFP4 weight the exporter runs, in order:
- GPTQ-OBS one-shot rounding — block-wise error propagation along the group-quant structure using the calibration Hessian. Closed-form, not iterative. Handles cross-column activation coupling.
- Closed-form per-group scale sweep — for each 16-weight NVFP4 group, enumerate
grid=32 candidate scales spanning [0.5·s₀, 1.5·s₀], round each weight to its nearest codebook neighbor at every candidate scale, pick the (scale, rounding-set) configuration minimizing activation-weighted per-group MSE sum_j a_j² · (w_orig,j - w_q,j)². Improve-or-keep gate against the post-GPTQ weight.
Measured per-Linear output-MSE vs RTN baseline (geomean):
| Pipeline variant | out_mse ratio vs RTN |
|---|
| RTN (no passes) | 1.00 |
| GPTQ only | 0.41 |
| GPTQ + act_round polish (prior pipeline) | 0.99 (act_round undid GPTQ) |
| scale_sweep only | 0.33 |
| GPTQ + scale_sweep (this artifact) | 0.33 |
AWQ's γ-fold is not applied. On NVFP4's 16-channel groups, AWQ's per-channel rescaling pushes mixed-scale values into the same group and inflates per-group quant noise rather than reducing it.
Pareto curve (body allocation)
The allocator sweeps a range of bit targets and reports predicted Δloss (Fisher-weighted MSE proxy). Target 4.75 bpp was chosen; the allocator suggested 4.5 bpp as the knee.
| Target bpp | NVFP4 | MXFP8_E4M3 | BF16 | Δloss (pred) |
|---|
| 4.50 | 246 | 1 | 31 | 1.148e+01 |
| 4.60 | 183 | 24 | 71 | 1.040e+01 |
| 4.70 | 140 | 40 | 98 | 1.009e+01 |
| 4.75 | 121 | 48 | 109 | 9.969e+00 |
| 4.85 | 94 | 39 | 145 | 9.795e+00 |
| 5.00 | 102 | 42 | 134 | 9.415e+00 |
| 5.25 | 99 | 48 | 131 | 8.646e+00 |
| 5.50 | 100 | 46 | 132 | 7.885e+00 |
| 6.00 | 87 | 38 | 153 | 7.408e+00 |
| 7.00 | 69 | 35 | 174 | 5.683e+00 |
| 8.25 | 54 | 30 | 194 | 3.524e+00 |
Which layers are quantized
Text body (DeltaNet linear-attention + dense MoE, 40 layers)
- Full attention Linears (
q_proj, k_proj, v_proj, o_proj): mixed NVFP4 / MXFP8 / BF16 per-Linear by sensitivity
- DeltaNet linear-attention Linears (
in_proj_qkv, in_proj_z, in_proj_a, in_proj_b, out_proj): same — note in_proj_a / in_proj_b variants were ineligible for MXFP8 due to shape constraints and resolved to NVFP4 or BF16
- MoE experts (
gate_up_proj, down_proj): per-expert NVFP4 (packed_moe_per_expert) with joint per-tensor scale across the gate_up pair so vLLM FusedMoE loads them (80 per-expert entries)
- Shared expert MLP: same per-Linear policy
- Router (
mlp.gate): always BF16 (tiny, sensitive)
Multi-token-prediction (MTP) head
Speculative-decoding head (1 layer) + its own MoE block quantized under the same per-Linear policy (3 089 MTP tensors merged at export). --speculative-config method=mtp drafts at the same precision as the body.
Visual encoder (110 Linears — BF16)
All 110 visual Linears are kept at BF16 via --visual-format=BF16. This is a deliberate conservative choice — the visual tower is left at full precision rather than running sensitivity-driven allocation on it. Disk overhead vs a quantized visual tower is small (< 2 GB) and visual quality is fully preserved.
Passthrough (unquantized)
lm_head — kept at BF16. vLLM's ParallelLMHead only accepts a single weight parameter; the compressed-tensors runtime rejects a compressed lm_head with KeyError: lm_head.input_global_scale. This is a vLLM runtime limitation, not a PrismaQuant design decision. The allocator would have chosen NVFP4 here.- RMSNorm weights (all layers + MTP + visual)
- All biases
embed_tokensmodel.visual.pos_embed (Parameter/Embedding, not an nn.Linear)
Serving (vLLM only)
This artifact is only runnable via vLLM's stock compressed-tensors support. vLLM 0.11+ is required.
1vllm serve cyburn/Qwen3.6-35B-A3B-Claude-4.7-Opus-Reasoning-Distilled-PrismaQuant-4.75bit-vllm \
2 --trust-remote-code \
3 --max-model-len 32768 \
4 --gpu-memory-utilization 0.90 \
5 --speculative-config '{"method":"mtp","num_speculative_tokens":3}'
Thinking / reasoning mode
This model was distilled from Claude 4.7 Opus chain-of-thought traces. Enable reasoning at inference time via the chat template:
1messages = [{"role": "user", "content": "..."}]
2text = tokenizer.apply_chat_template(
3 messages,
4 tokenize=False,
5 add_generation_prompt=True,
6 enable_thinking=True, # enables <think>...</think> CoT blocks
7)
Or pass /think at the start of your prompt in interactive use. The model works without thinking mode enabled — it will respond directly — but reasoning quality is higher with it on.
- FlashInfer NVFP4 attention is picked up automatically; set
VLLM_USE_FLASHINFER_NVFP4=1 to make the preference explicit.
- MTP speculative decoding at
n=3 is the measured optimum for this family (n=2 leaves ~10 % tok/s on the table, n=4 regresses).
- Visual inputs work via vLLM's standard
image-text-to-text chat API — no special flags.
Reproducing this artifact
Full pipeline is in the
PrismaQuant repo:
- Sensitivity probe — streaming per-shard empirical-Fisher trace (diagonal) across body + MTP Linears.
- Per-(Linear, format) cost measurement — for each Linear and each candidate format, the per-group RTN error weighted by cached input activations.
- Multi-choice knapsack allocator — picks one format per Linear minimizing total predicted Δloss under the bit budget. Target 4.75 bpp.
- Export — streams each body / MTP shard, applies GPTQ + activation-weighted scale sweep to its NVFP4 entries, writes the compressed-tensors format. Visual encoder stamped uniformly at BF16 via
--visual-format=BF16.
Known issues / limitations
- vLLM only at serve time. No transformers-runtime path for this precision mix today.
- lm_head stays BF16 because vLLM's
ParallelLMHead does not register the NVFP4/MXFP8 compressed-tensors schemes. Costs a small amount of disk vs full NVFP4.
- Visual encoder is BF16. If disk size is the primary constraint and visual quality is not critical, a variant with sensitivity-driven visual quantization (as in
rdtand/Qwen3.6-35B-A3B-PrismaQuant-4.75bit-vllm) could reduce the visual tower to ~22 % of its BF16 size.
- MTP n=4 regresses on this family. Stick to
n=3 unless you verify against the draft-head acceptance-rate trace.
Links
Citation
1@software{prismaquant2026,
2 title = {PrismaQuant: per-Linear sensitivity-driven mixed-precision
3 quantization for LLMs},
4 author = {Tand, Rob},
5 year = 2026,
6 url = {https://github.com/RobTand/prismaquant},
7}
