Qwen3.8-Flash-Next Coder for the Tenstorrent P150 (Strata tt backend)

Block-float weights that let Strata's tt backend run Qwen3.8-Flash-Next Coder on one Tenstorrent Blackhole P150 card:

• the routed experts of ISTA-DASLab/Qwen3.8-Flash-Next-GSQ-RCO-Coder-GGUF (IQ1_M), 48 layers x 256 experts x gate/up/down, re-quantized with GPTQ to Tenstorrent's BFP4 (bfp4_b) format, in the exact byte layout the P150 kernels read;
• the precision/placement plan bfp4-tiered-2.4e+10: which experts live in the card's DRAM and which in pinned host memory;
• the MTP draft layer's 512 routed experts in BFP4 and their routing profile (speculative decoding).

This is not a standalone model. It holds no tokenizer, attention, embedding or LM-head weights: those come from the GGUF that Strata's setup downloads. The files are only useful to the Strata tt backend (the fnx engine in Strata's tt/ folder) and are fetched by it at a pinned commit (tt/artifacts.json, tt/tools/tt_artifacts.py fetch).

Files (340 files, 59.8 GB)

Paths are relative to the backend's data folder (FNX_DATA).

How it was made

• Source: the GGUF shards of ISTA-DASLab/Qwen3.8-Flash-Next-GSQ-RCO-Coder-GGUF at revision 5348543e0147355ac9cbcb031184a3546350988e, IQ1_M (the revision Strata pins), a quantization of Qwen/Qwen3.8-Flash-Next that keeps 256 of its 512 routed experts.
• Calibration (fnx.quant.calib_data): 128 sequences x 2048 tokens (+16 held out), chat 48 (UltraChat 200k train_sft), code 48 (Python, C/C++/CUDA, TypeScript/JavaScript, Rust, Go source trees), instruct 16 (databricks-dolly-15k), docs 16 (Markdown documentation); disjoint from the quality-gate evaluation set.
• GPTQ-BFP4 (fnx.quant.pipeline quantize --weighted 0 --n-min 8192): a float32 PyTorch forward of the GGUF (dequantized weights) over the calibration set, layer by layer on one RTX 3090 (62 minutes for 48 layers); per expert input Hessians of its routed rows (gate/up share one; a layer-mean prior up to 8192 rows for rarely routed experts), then GPTQ with act-order, 1% damping and a min-MSE shared exponent per 16-value block (clips -1..2), emitted as bit-exact tt-metal bfp4_b blocks.
• Plan (fnx.quant.alloc --tiered 2.4e10): every expert BFP4; the 8680 most-routed experts (calibration routing counts, 85.9% of all routings, 24.0 GB) are resident in the P150's DRAM, the other 3608 are read from pinned host memory over PCIe by the MoE kernels.
• MTP drafter: the MTP layer of Qwen/Qwen3.8-Flash-Next as Strata packs it (Q2_0 experts), converted to BFP4 with a per-block min-MSE exponent (fnx.quant.methods.search); the routing profile comes from draft rounds on calibration text on the P150 (fnx/tests/test_mtp.py profile).
• Full model (IQ3_S) card residents (fnx.quant.residents): the same calibration set through a float32 forward of ISTA-DASLab/Qwen3.8-Flash-Next-GSQ-RCO-GGUF IQ3_S (revision ed59f92082b1e93c0e96d60a8b11aab089b52f09, all 512 experts), Hessians of the 8,823 most-routed experts only (the ones the card holds; the rest run on the CPU from the GGUF as they are), the same prior and damping; GPTQ with the block exponent fixed at each 16-column block's first column and whole blocks in act-order (fnx.quant.methods.gptq_rows); 30 minutes on one RTX 3090. Quality gate on the P150 (cpu MoE with these residents): KL .0279 (.0437 with min-MSE BFP4 of the GGUF values), top-1 .942 (.925), perplexity -0.66% (-2.21%), GSM8K-50 45/50 (45/50); same speed.

tt/tools/tt_artifacts.py regenerate in Strata reruns these steps (CUDA GPU with >= 16 GB).

Quality and speed (P150, Strata tt backend, default configuration)

Quality gate against the published GGUF evaluated in float32 (teacher-forced chat/code/doc records + GSM8K), with these BFP4 experts on the device (gated 2026-10-07 on the 22.0 GB placement of the same rows; the placement does not change the numerics):

Decode speed on one P150, traced, chat / code / doc prompts: 30.9 / 29.2 / 26.6 tok/s plain, 46.9 / 50.9 / 50.0 tok/s with MTP speculative decoding (2026-10-08). Context up to 32,768 tokens; text only.

License

These weights are a derivative of Qwen/Qwen3.8-Flash-Next and are distributed under the Qwen Community License 1.0 (the LICENSE file here, copied unchanged from the base model repository). Its condition 1 requires that copyright and permission notice in all copies. Its condition 2 applies to every downstream user: a licensee (or affiliate) that runs a Model as a Service or AI Work Assistant business must obtain a separate license from Qwen before using the weights or derivatives for any commercial purpose (internal use that exposes nothing to third parties excepted).

The routed experts were re-quantized from ISTA-DASLab's Qwen3.8-Flash-Next-GSQ-RCO-Coder-GGUF (GSQ-RCO expert pruning and quantization by IST Austria's DASLab, published under Apache-2.0); credit for the expert selection and the IQ1_M quantization goes to them. The MTP drafter weights come from the Qwen base checkpoint.