Digisensus

Qwen3.8-27B-Digisensus-NVFP4-RTX5090-256K

README

License: apache-2.0

How close is it to the original model?

We measure how far the quantized model's next-token predictions are from the original bf16 Qwen3.8-27B ("stock"): the KL divergence, averaged over every token of 125 held-out prompts (prose, tool calls, JSON, reasoning, code). Lower is closer to stock. "Top-1" is how often the quantized model picks the same next token as stock.

For comparison we include another NVFP4 quant of the same model that was also adjusted for the RTX 5090. Both use the same amount of GPU memory.

Table with columns: model, KL vs stock (lower is better), top-1 agreement with stock
model	KL vs stock (lower is better)	top-1 agreement with stock
another 5090-adjusted NVFP4 quant	0.01520	95.27%
Qwen3.8-27B-Digisensus	0.01148	96.07%

Digisensus is 24.5% closer to stock (95% confidence interval 21.8%–26.9%). By task: prose 26% closer, reasoning 26%, code 22%, tool calls 19%.

Thinking on, 24 real requests (JSON scoring, tool agents, reasoning): Digisensus wrote 8% fewer reasoning tokens, hit the token limit 2 times instead of 5 (of 48 runs), and gave 46 valid answers instead of 43. MTP accepted 67.9% of draft tokens instead of 65.9%.

How it was made

We measured, for every part of the network, how much quality it loses at FP8 and at NVFP4, then gave more bits to the parts that lose the most and fewer to the parts that lose almost nothing — at the same total memory:

embeddings: bf16 → INT8 per row. Costs almost nothing, frees 1.18 GiB. This paid for everything else.
MLP layers 28–35 and 48–55: NVFP4 → FP8. These 4-bit MLPs lost the most.
MLP layers 56–59: FP8 → NVFP4. They lose less than 48–55.
layer-63 attention o_proj: bf16.
Gated-DeltaNet (linear attention) projections stay at FP8: quantizing them harder hurts long reasoning. MTP head bf16. Vision tower bf16 (not used with --language-model-only).

Table with columns: part, format
part	format
MLP layers 0–27, 36–47, 56–59	NVFP4 W4A4 (groups of 16, FP8 scales), GPTQ
MLP layers 28–35, 48–55, 60–63 (`down` of 28–31: NVFP4)	FP8 W8A8
attention q/k/v/o, GDN in_proj_qkv/z/out_proj, lm_head	FP8 W8A8 (layer-63 o_proj: bf16)
embed_tokens	INT8 per row, weight only
KV cache	served as NVFP4 (FP8 scales in the checkpoint)
MTP head (`model_mtp.safetensors`)	bf16

The full method, every script and all raw numbers: GitHub — Digisensus/Qwen3.8-27B-Digisensus-NVFP4-RTX5090-256K.

How to run it

You need the vLLM v0.28.0 image and patches from the GitHub repo. Stock vLLM cannot load these weights (INT8 embeddings) and cannot run the NVFP4 KV cache on an RTX 5090. The repo also fixes stop strings firing inside thinking, leading newlines in answers, and JSON quote truncation.

bash

git clone https://github.com/Digisensus/Qwen3.8-27B-Digisensus-NVFP4-RTX5090-256K.gitcd Qwen3.8-27B-Digisensus-NVFP4-RTX5090-256K./image/build.shhf download Digisensus/Qwen3.8-27B-Digisensus-NVFP4-RTX5090-256K --local-dir models/Qwen3.8-27B-Digisensus-NVFP4-RTX5090-256Kcd compose && cp .env.example .env && docker compose up -d

Main flags:

text

--max-model-len 262144 --kv-cache-dtype nvfp4 --kv-cache-memory-bytes 5694957226 --gpu-memory-utilization 0.94 --speculative-config '{"method":"qwen3_5_mtp","num_speculative_tokens":3}' --language-model-only --reasoning-parser qwen3 --tool-call-parser qwen3_xml

Sampling for tool calls and JSON: presence_penalty=0 (Qwen's suggested 1.5 makes the model leave repeated tool arguments empty), temperature 0.7 or lower.

Speed (one RTX 5090, 480 W and 600 W, the compose config above)

Prefill (reading the prompt), one request at a time, random-token prompts:

Table with columns: context, prefill tokens/s, time to first token
context	prefill tokens/s	time to first token
1k	4,475	0.23 s
8k	8,517	0.96 s
16k	10,703	1.5 s
32k	6,494	5.1 s
64k	4,986	13 s
128k	3,347	39 s

Decode (writing the answer) on real text: a long document of real conversations plus "summarise", 384 output tokens, thinking off, temperature 0.7 (quantization/decode_ctx.py). Random-token prompts are not used for decode because MTP speed depends on how predictable the text is. Prefix caching was switched off for these runs so nothing is reused between requests. Same engine, same test, only the GPU power limit changed. Single-user numbers are the engine's own counters (tokens / decode-step time; two runs each agree within noise).

Table with columns: context, 1 user: tokens/s, 480 W, 1 user: tokens/s, 600 W, step time 480 / 600 W, users at once, all users together: tokens/s, 480 W, 600 W, 600 W vs 480 W
context	1 user: tokens/s, 480 W	1 user: tokens/s, 600 W	step time 480 / 600 W	users at once	all users together: tokens/s, 480 W	600 W	600 W vs 480 W
1k	113	118	23.0 / 23.1 ms	8	692	807	+17%
8k	119	120

Peak power draw: 495 W at the 480 W limit, 610 W at the 600 W limit.

What the table says:

One user needs no power. At 480 W a single request runs at the same speed as at 600 W at every context (the differences are run-to-run noise). A decode step for one user is limited by reading the 21 GB of weights, not by compute, so the extra 120 W buys nothing.
Many users at short context are where 600 W helps: 8 users at 1k get 807 tokens/s at 600 W vs 692 at 480 W (+17%). At 8k and beyond the gain is 4–9%.
Decode barely depends on context because 48 of the 64 layers are Gated-DeltaNet (linear attention), whose cost per token does not grow with context; only the 16 full-attention layers read the (NVFP4) KV cache. The step time grows from 23.0 ms at 1k to 30.0 ms at 258k (+30%); MTP then turns each step into 2.6–3.3 tokens.
"All users together" = all output tokens divided by the time during which at least one user was receiving tokens. Per-user speed drops at long context because the engine mixes other users' prompt reading into the decode steps. The number of users at once is limited by KV memory: min(max-num-seqs, KV pool / (context + output)).

More than 8 users (480 W, high-concurrency profile)

The config above allows 8 requests at once and the full 256K context. To see whether more users add throughput we ran a second profile at 480 W: --max-num-seqs 32, --max-model-len 139264 (136K), KV pool sized by vLLM (198,767 tokens), --gpu-memory-utilization 0.90 (compose/docker-compose.high-concurrency.example.yaml). Why not 64 users or 256K here: every sequence reserves a Gated-DeltaNet recurrent state (about 75 MB across the 48 GDN layers) up front, so 64 sequences cost 4.8 GB and 32 sequences no longer leave room for one 256K request.

Table with columns: context, users at once, per user: tokens/s, all users together: tokens/s
context	users at once	per user: tokens/s	all users together: tokens/s
1k	8	111	777
1k	16	93	752
1k	32	85	633
8k	8	73

More users do not add throughput: with MTP every step checks 4 draft tokens per user, so the batch grows fast and past about 8 users the total goes down, not up. On this card and this quant, 8 users is the sweet spot; more users only spread the same throughput thinner.

Calibration data

About 641k tokens, rendered with the model's own chat template: 50% private Digisensus.com test traffic (contact-centre QA, audit, PII and scoring prompts, tool-calling agents, with the model's own answers; not released), 18% NousResearch/hermes-function-calling-v1 tool calls (Apache-2.0), 12% hermes json-mode (Apache-2.0), 20% HuggingFaceH4/ultrachat_200k (MIT). Calibration only looks at activations; nothing was trained.

Limits

Needs the patched vLLM from the GitHub repo.
Tuned and judged for agentic use, structured output and reasoning.
Long-context retrieval tests (needle, RULER) at 256K with the NVFP4 KV cache are not published yet.

License

Apache-2.0, derived from Qwen/Qwen3.8-27B. The serving stack builds on vLLM (Apache-2.0); third-party notices are in the GitHub repo's NOTICE file.

Available on FriendliAI

Dedicated Endpoints

Run this model inference on single tenant GPU with unmatched speed and reliability at scale.

Learn more

Model Details

Model Provider

Digisensus

Model Tree

Base

Qwen/Qwen3.8-27B

Quantized

this model

Input Modalities

TextImageVideo

Output Modalities

Text

Supported Functionality