logic65/Whittle-Qwen-3.8-35B-A3B
One-line: First Whittle release where the 10B-param n-gram memory is load-bearing (zeroing it measurably hurts the model) — 35.1B total / ~3B active MoE in the Qwen3.8-Flash-Next (qwen4_exp) format, runs on stock llama.cpp, distilled from Qwen3.8-27B thinking traces.
⚠ no published standard benchmarks — agent-loop suitability unknown. Card ships private probes only (GSM8K subset, a 60-problem MATH-train maths probe, a private long-context PR-diff gate); self-described as “research preview, not a finished distillation”. Repo has full weights + eval logs; GGUFs in sibling repo. Released 2026-09-19 (v4.4→tbl1→lw2 lineage; lw2 is root).
Coding benchmarks
| Benchmark | Score | Source |
|---|---|---|
| HumanEval | — | — |
| HumanEval+ | — | — |
| MBPP | — | — |
| LiveCodeBench | — | — |
| Aider polyglot | — | — |
| SWE-bench Verified | — | — |
| SWE-bench Multilingual | — | — |
| Long-context reading gate (private: 6 levels 4k→75k, real PR diff) | 30/36 | lab (lw2; production Qwen3.6-35B-A3B scores 18/36) |
Agentic / tool-use benchmarks
| Benchmark | Score | Source |
|---|---|---|
| BFCL | — | — |
| τ-Bench | — | — |
| ToolACE | — | — |
| GAIA | — | — |
Reasoning (context only)
| Benchmark | Score | Source |
|---|---|---|
| MMLU / MMLU-Pro | — | — |
| GPQA Diamond | — | — |
| MATH (private 60-problem probe, levels 2–4) | 44/60 | lab (lw2; tbl1 46/60) |
| GSM8K (private 50-problem subset) | 41/50 | lab (lw2; tbl1 44/50) |
Efficiency
- Active params: ~3B per token (8 of 180 routed experts + shared expert; 10B memory is a lookup, not compute) — total 35.1B = 25.1B body + 10.0B n-gram memory
- Context length: 262k positions (as parent Qwen3.8-Flash-Next); tested to 75k (PoSE-trained at offsets up to 131k)
- VRAM (fp16 rough): body-only fp16 ≈ 50GB — not single-GPU. The practical path is the Q8_0 GGUF with memory in system RAM: ~25GB GPU footprint (27B-class body), 10.5GB RAM for the memory table, generation speed of a 3B model (author-verified on 3× RTX 3060). Math caveat:
bf16-tbl1/column scores 2 points higher on maths if that matters more than long-context.
What makes it notable
Novel post-training recipe (table-first warm-up + dependence loss that punishes the body for being as good without the memory; forward-KL distillation on 1,840 complete Qwen3.8-27B thinking traces; PoSE long-context training), and it is the first Whittle where the n-gram memory transfer from Qwen3.8-Flash-Next is actually load-bearing (memory gain +2.84 nats held-out code). Stop/loop battery is clean (24/24, 0 loops, 0 self-turn leaks) under the serving sampler. For an agent-loop buyer this is a watch item, not a drop-in: no SWE-bench/Aider/BFCL numbers, card warns about occasional invalid-JSON for bare structured output (mitigate with JSON-schema grammar at serve time), and thinking length is unbounded by default. Runs on stock llama.cpp with the qwen4exp architecture (update llama.cpp if unknown). Apache-2.0, weights + probes public.
See also
- welcome
- Source: https://huggingface.co/logic65/Whittle-Qwen-3.8-35B-A3B
- Family: Whittle-Next-27B-A3B · GGUFs: logic65/Whittle-Qwen-3.8-35B-A3B-GGUF