LibertAIDAI/DeepSeek-V4.1-Flash-1017-EXL3-2.9bpw

DeepSeek-V4.1-Flash-1017 · EXL3 2.9 bpw

TL;DR — EXL3 quantization of deepseek-ai/DeepSeek-V4.1-Flash-1017 at 2.90 bpw (196.3 GiB). On held-out text its perplexity is +1.8% (code) / +0.4% (prose) above the unquantized model. Tensor names, dtypes and per-tensor bitrates match Mia-AiLab/DeepSeek-V4.1-Flash-EXL3-2.9bpw tensor for tensor, so the MiaAI-Lab/DeepSeek-v4.1-Flash-EXL3-2x-DGX-Sparks serving kit's bundled K map describes it unchanged. The two Engram tables are not in this repo: take them from the 1017 base repo, because they changed in this release.

Chat template: this repo ships chat_template.jinja, a Jinja port of DeepSeek's 1017 encoding/encoding.py (from jayleaton's MIT template via a community PR by dpolist, amended for the 1017 encoder). It renders byte-identically to the reference on our 8,455-case test corpus. Runtimes that require a template file at startup need it (verified to render identically under Python Jinja and TensorFold's Zig engine); vLLM-based kits use their built-in DeepSeek encoder.

Agentic use / reasoning effort: with tools in the conversation, each turn's reasoning stays in the prompt and is re-read by every later turn, so reasoning cost compounds. On our benchmark of the original 1017 weights (3,738 jobs: agentic coding, tau-bench, LiveCodeBench, AIME, chat), reasoning_effort high (75, the default) matched max (100) on quality at 0.55–0.8× the cost, and low (50) matched too at 0.4–0.6× (results in discussion #4). Avoid max for agent loops unless you need the headroom.

Base model	deepseek-ai/DeepSeek-V4.1-Flash-1017 (MIT)
Format	EXL3 trellis, mul1 codebook, output channel scales always
Bitrate	2.90 bpw over the quantized trunk weights, head 6 bpw
Size	210.7 GB (196.3 GiB), 43 safetensors shards
Perplexity vs unquantized	code 1.9094 vs 1.8749, prose 6.4520 vs 6.4277
Engram tables	not included; shards 47 + 48 of the base repo
Vision tower	included, BF16
MTP / DSpark drafter	included
Other releases	LibertAIDAI/DeepSeek-V4.1-Flash-1017-EXL3-3.5bpw · LibertAIDAI/DeepSeek-V4.1-Flash-1017-NVFP4

Tensor groups

Group	Tensors	Stored as	Notes
Routed experts	layers.N.ffn.experts.E.w1/w2/w3, 384 per layer	EXL3 3 bpw; layers 18-22: 2 bpw	one K per layer for all 1,152 matrices (the kit's MoE loader needs that)
Shared expert	layers.N.ffn.shared_experts.w1/w2/w3	EXL3 5 bpw on layers 0-10, 30-39; 4 bpw on 11-28	layer 29: w1/w3 4, w2 5
Attention	wq_a, wq_b, wkv, wo_a.slice.0..7, wo_b	EXL3 5 bpw	layer 0 wq_a / wkv: 6 bpw; wo_a stored as 8 per-group slices
KV compressor	attn.compressor.wkv / wgate (layers 2, 8, 14, 20)	EXL3 5 bpw	compressor.norm BF16
Indexer	attn.indexer.wq_b (layers 2, 8, 14, 20, 24, 28, 32, 36); indexer.wk (2, 8, 14, 20)	EXL3 5 bpw; wk 8 bpw	weights_proj FP16, k_norm BF16
Engram projection	layers.1/14.engram.wkv	EXL3 5 bpw (layer 1), 4 bpw (layer 14)	q_weight / k_weight FP32
Engram n-gram tables	layers.1/14.engram.embed	not in this repo	read unquantized from the base checkpoint, see below
Output head	head	EXL3 6 bpw
MTP / DSpark drafter, routed experts	mtp.0-2.ffn.experts.*, 128 per block	EXL3 4 bpw	matches the kit's mtp_bits 4
MTP / DSpark drafter, dense	attention, shared expert, mtp.0.main_proj	EXL3 6 bpw	4 bpw in the earlier V4.1 release
Token embedding	embed	BF16
Router	ffn.gate.weight / gate.bias / gate.bias_vl	FP16 / FP16 / FP32	gate.bias_vl = routing bias for image tokens
mHC streams, attention sinks	hc_attn_*, hc_ffn_*, attn.attn_sink	FP32
Norms	all RMSNorm weights	BF16
DSpark heads	mtp.2.markov_head.*, mtp.2.confidence_head.proj	BF16 / FP16
Vision tower + aligner	vision.*, aligner.*, image_start/end/newline	BF16	copied from the base checkpoint unchanged

Every EXL3 matrix is stored as trellis (int16) + suh / svh (fp16) + mul1 (int32 codebook marker).

Perplexity

Two held-out sets of unpublished text, 32 sequences x 2048 tokens each, tokens 2-2048 of every sequence scored (65,504 tokens per set). Every row uses the same full-sequence forward (see How it was made).

Model	Size	code	prose
Unquantized (base FP8 / MXFP4 weights, dequantized)	—	1.8749	6.4277
EXL3 2.9 bpw (this repo)	196.3 GiB	1.9094	6.4520
EXL3 3.5 bpw	239.9 GiB	1.8895	6.4341
This repo with Engram switched off (sanity check)		4.1648	9.0627

Engram tables (required)

DeepSeek-V4.1 adds two Engram n-gram lookup tables (layers.1.engram.embed, layers.14.engram.embed: FP8 with E8M0 block scales, ~95 GiB each). They are used unquantized and come from the base checkpoint, like in the existing V4.1 EXL3 serving stacks. The 1017 tables are new: the shards from the older deepseek-ai/DeepSeek-V4.1-Flash must not be reused.

File from deepseek-ai/DeepSeek-V4.1-Flash-1017	Bytes	sha256
model-00047-of-00048.safetensors	101,535,150,936	76a05106ece259be83d1bfea0dcb54f23377864ee60fd321690f35e6698a2af5
model-00048-of-00048.safetensors	101,537,926,640	9dc51433e195bc78b5ad846b68643cc0351d3b174e5e4309222d3b7402c83afd
model.safetensors.index.json, config.json		needed for engram_layer_ids / engram_num_embeddings

hf download deepseek-ai/DeepSeek-V4.1-Flash-1017 \
  --include model-00047-of-00048.safetensors model-00048-of-00048.safetensors model.safetensors.index.json config.json \
  --local-dir ./engram-src-1017
sha256sum engram-src-1017/model-0004[78]-of-00048.safetensors

Serving

The checkpoint follows the layout of Mia-AiLab/DeepSeek-V4.1-Flash-EXL3-2.9bpw, so it targets the vLLM-based MiaAI-Lab/DeepSeek-v4.1-Flash-EXL3-2x-DGX-Sparks kit (2x DGX Spark, TP=2). The settings below come from the kit's start.sh / download.sh (commit 6f7d159). Upstream exllamav3 has no DeepseekV41 model class yet, so it cannot load this checkpoint directly.

Fresh install (on the head node, from the kit checkout):

git clone https://github.com/MiaAI-Lab/DeepSeek-v4.1-Flash-EXL3-2x-DGX-Sparks
cd DeepSeek-v4.1-Flash-EXL3-2x-DGX-Sparks
cp .env.example .env
cat >> .env <<'EOF'
HF_MODEL_REPO=LibertAIDAI/DeepSeek-V4.1-Flash-1017-EXL3-2.9bpw
MODEL_HOST=$PWD/model-1017-2.9bpw
EXPECTED_SHARDS=43
HF_ENGRAM_REPO=deepseek-ai/DeepSeek-V4.1-Flash-1017
ENGRAM_DIR=./engram-src-1017
EOF
./download.sh   # optional: stage both downloads first
./start.sh

.env is sourced by bash, so $PWD expands to the kit directory when the scripts are run from it.

The kit's bundled files/exl3_k_map.json already matches this checkpoint (this repo's exl3_k_map.json has the same values for every key), so KMAP_HOST can stay at its default.

Switching an existing V4.1 install — the kit reuses anything already on disk, so make sure nothing old gets picked up:

What	Why	Do
MODEL_HOST	weights are only downloaded when fewer than EXPECTED_SHARDS shards are present, so an old 39-shard tree is served as-is	point it at a new directory (as above)
ENGRAM_DIR	the Engram fetch only fills missing files, so old shards 47 / 48 would stay	new directory (as above); check the sha256
Packed Engram (./start.sh pack, ~/dsv41-engram on both nodes)	used automatically whenever engram-l1-r*of2.bin exists	delete it or set HEAD_PACKED_DIR / WORKER_PACKED_DIR to new paths, then re-run ./start.sh pack
WEIGHT_SYNC=rsync worker copies	skipped when the .dsv41-exl3-synced / .dsv41-engram-synced markers exist	new WORKER_MODEL_DIR / WORKER_ENGRAM_DIR

How it was made

Step	Details
Quantizer	exllamav3 dev branch (ba2feb2, 1.6.2), its own convert.py pipeline: per-module Hessian capture, LDLQ + trellis encoding
Architecture	a DeepseekV41 model definition written for this release against DeepSeek's reference inference/model.py / inference/engram.py: lagged-pre mHC residual streams, compress ratios 0 / 2 / 1 with cross-layer KV / indexer / candidate sharing, Engram lookups, sqrt-softplus routing, the reference FP8 / FP4 activation quantization inside attention
Forward check	the same forward on the unquantized weights gives the perplexity in the table above (code 1.8749 / prose 6.4277)
Calibration	exllamav3 default mix (C4, code, multilingual, technical, Wikipedia, TinyStories, random-token rows), 250 rows x 2048 tokens
MoE capture	every routed expert sees every calibration token (exllamav3's default for MoE models)
MTP / DSpark	quantized without calibration data
Bitrate plan	explicit per-tensor recipe, integer bitrates only, one K per routed-expert layer
Validation	held-out perplexity, Engram on / off, chat smoke test with the base repo's encoding/encoding.py (below)

Smoke test outputs (greedy decoding)

chat: What is 17*19? Reply with the integer only.

323<｜end▁of▁sentence｜>

chat: Explain in three sentences why the sky is blue.

Sunlight contains all colors, but air molecules scatter short blue wavelengths much more than longer red ones—a process called Rayleigh scattering. This scattered blue light reaches your eyes from all directions, making the sky appear blue. At sunrise and sunset, light travels through more atmosphere, so blue is scattered away and reds and oranges dominate.<｜end▁of▁sentence｜>

thinking: A bat and a ball cost $1.10 in total. The bat costs $1.00 more than the ball. How much does the ball cost?

We need answer classic. Need be careful: ball 0.05, bat 1.05. Explain concise.</think>The ball costs **$0.05** (5 cents).

The bat costs $1.05, so together they total $1.10.<｜end▁of▁sentence｜>

Files

File	Content
model-000NN-of-00043.safetensors, model.safetensors.index.json	weights
config.json	base config plus quantization_config (same keys as the earlier V4.1 EXL3 release) and engram_table_dir: null
exl3_k_map.json	per-tensor K in the kit's files/exl3_k_map.json schema, generated from these shards
quantization_config.json	per-tensor storage map written by exllamav3
tokenizer.json, tokenizer_config.json, LICENSE, encoding/	copied from the base repo (encoding/ is DeepSeek's prompt-format reference)

Credits

• DeepSeek for deepseek-ai/DeepSeek-V4.1-Flash-1017 (MIT).
• turboderp for exllamav3 (MIT): quantizer, kernels and conversion pipeline.
• Mia-AiLab for Mia-AiLab/DeepSeek-V4.1-Flash-EXL3-2.9bpw and the MiaAI-Lab/DeepSeek-v4.1-Flash-EXL3-2x-DGX-Sparks kit, whose tensor layout and K map this release follows.

License

MIT, same as the base model.