DeepSeek V4.1 Flash

GSQ–RCO IQ2_XS · AtomicChat · Measured configuration

83.19 / 100 Mango weighted score · 71.43 min matched workload · 12/12 mechanical agent passes

AtomicChat's refusal-direction-edited IQ2_XS GSQ–RCO package of DeepSeek V4.1 Flash 1017. The four main shards are preserved here; the native Mango setup also used the exact Engram file and Q8 DSpark companion linked below.

Benchmark results · Files · Download · Runtime · Provenance · Full collection

Join Infinity on Discord

Get model support, request models and quants, discuss benchmarks, and follow new releases and model improvements.

Join the Infinity community

All model tests, per-question scores, experiment history and scoring methods

Qwen effort study — separate full-404 cohort: medium 82.87/100 versus xhigh 84.41/100; paired difference +1.54 points (95% interval [-1.71, +4.82]). 404 collected questions, 403 valid questions graded per effort; one invalid reference annotation excluded symmetrically. The historical 202-question cross-model comparison below is unchanged. Study, timing and blind subjective results.

Preserved and documented by isilicon / Mango. The original model, modification and quantization remain the upstream creators' work. This repository adds attributed preservation and exact-configuration evaluation, with the selected source bytes unchanged.

At a glance

GB is decimal and GiB is binary. Payload size is disk/storage size, not a claim that the model fits in the same amount of VRAM. Runtime, KV cache, working memory and companions add their own requirements.

Mango evaluation

Cybersecurity is its strongest comparative point: 87.22 across three equally averaged cyber tasks. It passed 17/20 local coding tests, 28/40 function-call checks and all 12 mechanical agent cases.

Its 83.19 weighted score is close to both Flash Next packages. The recorded workload took 71.43 minutes, approximately twice the previous Flash Next run. The quality difference from Flash Next is inconclusive; the extra time is a measured cost.

Domain scores equally average their component task means. The overall score applies the weights above. EQ is continuous; the 12-case mechanical agent pass rate is only one component of the agentic domain.

Task-level quality and uncertainty

These are small released subsets and local graders. HumanEval/BFCL scores are not the complete official benchmark pipelines. Agent checks cover completion, evidence, permissions and applicable pricing totals; other answer semantics remain unjudged.

Recorded performance

Workload time excludes loading, restarts, wipe downtime and idle gaps. Latency/TTFT summaries cover suite requests only. The reported token rate divides engine usage counts by request time; it is not native decode throughput and token accounting differs across engines. No concurrent-serving result is implied.

Matched cohort comparison

The old mixed GLM row is a historical, internally capped baseline whose exact weights are unavailable. Other rows also differ in quant, engine, context and effort. Point ranks do not settle the quality ordering.

Paired score differences

Intervals use 10,000 paired task-stratified bootstrap resamples, seed 41, with fixed task/domain weights. An interval containing zero does not establish a quality winner. See methodology for scope and definitions.

Download the current full numerical report and methodology. The original compact evaluation is retained as the earlier preservation snapshot; its model score is unchanged, while the current report includes the expanded cohort. The numerical report contains aggregate statistics and hashes; no prompts or generated answers are published.

Runtime and compatibility

This was tested through the Strata DeepSeek family, not as a generic standalone llama.cpp load. Restore the compatible loader, conversion/pack sidecars and required Engram layout. A collection of GGUF shards alone is not the full native runtime.

32,768-token native test context; high effort; arch=deepseek41, reasoning_budget_tokens=0, answer_room=false.

The local evaluation covers text/tool tasks. It does not establish the full advertised context ceiling, image/audio quality or every optional speculative-decoding path. Publisher capabilities and Mango-verified behavior are separate scopes.

Companion artifacts

The selected payloads of this model and its linked companions total 273.80 GB before runtime/cache files. That is a download/storage total, not a RAM/VRAM fit guarantee.

Files and integrity

Full paths, byte sizes and complete SHA-256/Git blob identities are in SOURCE_MANIFEST.json. Available original README/license/notice files are also preserved. MIRROR_STATUS.json records the earlier source-file verification; MODEL_CARD_STATUS.json records this documentation refresh separately.

Download and use

The commands pin the already-verified mirror weight revision. Documentation can advance on main without changing that preserved weight snapshot. Keep original subfolders and all shards.

python -m pip install -U huggingface_hub

hf download isilicon/mango-deepseek41-gsq-iq2xs-uncensored \
  DeepSeek-V4.1-Flash-1017-GSQ-RCO-IQ2_XS-00001-of-00004.gguf \
  DeepSeek-V4.1-Flash-1017-GSQ-RCO-IQ2_XS-00002-of-00004.gguf \
  DeepSeek-V4.1-Flash-1017-GSQ-RCO-IQ2_XS-00003-of-00004.gguf \
  DeepSeek-V4.1-Flash-1017-GSQ-RCO-IQ2_XS-00004-of-00004.gguf \
  --revision ec435fd44c1c5eb5cdb61b97cad96b218a316032 \
  --local-dir ./weights/deepseek41-gsq-iq2xs-uncensored

hf download isilicon/mango-deepseek41-gsq-engram \
  DeepSeek-V4.1-Flash-Engram-IQ4_NL.gguf \
  convert_ds41_upstream.py \
  --revision 23749573fdb80daf88dd108d93843b6b2ec9799f \
  --local-dir ./weights/deepseek41-gsq-engram

hf download isilicon/mango-deepseek41-dspark-q8 \
  dspark-DeepSeek-V4.1-Flash-1017-Q8_0.gguf \
  --revision f0721c818cc1716d63bf9d183af67c3e27080f16 \
  --local-dir ./weights/deepseek41-dspark-q8

Load example and setup scope

The recorded native Strata configuration must select the DeepSeek family and disable its artificial reasoning/answer-room limits:

{"arch": "deepseek41", "reasoning_budget_tokens": 0, "answer_room": false}

These are relevant configuration fields, not a complete loader configuration. Restore the native GSQ conversion/pack sidecars, exact Engram link and DSpark integration using the pinned publisher guide. Mango verified this engine configuration; this card does not claim a generic llama.cpp load.

Provenance and license

The pinned publisher card credits DeepSeek for the base, ISTA-DASLab for GSQ/RCO quantization, and AtomicChat for the refusal-direction edit. Its publisher-side capability and refusal measurements are distinct from Mango's task results.

Retained source terms: LICENSE.

Original notices, base-model conditions and the publisher's stated limitations still apply. Downloadable preservation does not add new license rights. Source-described refusal modification and publisher-side benchmark claims have not been independently re-measured by this Mango task cohort.

Original publisher documentation

The unchanged upstream README and pinned source card retain the original usage guidance, technical discussion, attribution, caveats and citations.

Read the complete original publisher card

Source reference. The following is the original publisher's documentation. Its other quants, full-model files, context claims and benchmark tables can exceed the selection in this mirror. Relative asset/file URLs are qualified to the pinned original source; publisher measurements remain separate from Mango measurements.

DeepSeek-V4.1-Flash-1017 Without Refusals · GSQ-RCO

ISTA-DASLab's GSQ-RCO quantizations of DeepSeek-V4.1-Flash-1017 (763B, 384 experts) with DeepSeek's refusal direction projected out of the dense residual writers — the same edit as AtomicChat/DeepSeek-V4.1-Flash-1017-Uncensored-GGUF, measured on every rung.

• Built by the team behind Atomic Chat, the app that runs open-weight models locally and offline.

---

• These are ISTA-DASLab/DeepSeek-V4.1-Flash-1017-GSQ-RCO-GGUF's GSQ-RCO backbone shards (revision cb02422c900d) with only the residual-writing dense tensors replaced: attention output (attn_output_b), the shared-expert down projection (ffn_down_shexp) and the Engram value projection (engram_wkv). The routed experts are left exactly as ISTA shipped them, so GSQ's learned low-bit rounding of the experts is preserved byte for byte, and each file stays the same size and runs wherever the original runs.
• The replaced tensors carry the one-direction refusal edit (strength 2.0) baked into DeepSeek's weights at full precision and quantized to ISTA's exact per-tensor type.
• Runs on a single 12 GB GPU with Strata: the backbone sits in system memory and the shared Engram table streams from an SSD, so a 763B model fits a consumer machine.

Not every rung cleared every gate; such a rung is published as measured, with its numbers in the table and the metrics repo.

What changed and what did not

Both files served on Strata, same prompts and template; "→" reads the stock GSQ-RCO rung → this one.

KL is the edit's distribution shift from the stock rung, from Strata's teacher-forced score mode: a top-64 lower bound, so the true KL is a little higher. Measurements, prompts and gates: AtomicChat/DeepSeek-V4.1-Flash-1017-Uncensored-GGUF-metrics (ablit/gsq/).

Running it

Download one rung and ISTA-DASLab's shared Engram file, then serve with Strata:

hf download AtomicChat/DeepSeek-V4.1-Flash-1017-Uncensored-GSQ-RCO-GGUF --include "*IQ2_XXS*" --local-dir ds-unc
hf download ISTA-DASLab/DeepSeek-V4.1-Flash-1017-GSQ-RCO-GGUF --include "*Engram*" --local-dir ds-unc   # the 110 GB lookup table, once
strata launch --model-path ds-unc/DeepSeek-V4.1-Flash-1017-GSQ-RCO-IQ2_XXS-00001-of-00004.gguf

The Engram file is identical to ISTA-DASLab's and is not re-hosted here; keep it on a fast SSD next to the shards. As of this writing these files run in Strata (DeepSeek-V4.1 with a separate Engram file is not yet in llama.cpp).

Limitations

• The refusal classifier reads the start of an answer, so a soft refusal later in the text can pass as an answer.
• The edit removes refusals across topics; it does not keep a narrower set of them.
• With thinking on, the edit plans an answer to a request the original declines, and the plan can be long; give such requests a larger token budget or a lower effort.

Responsible use

This model answers requests the original declines. You are responsible for how you use it and for following the laws that apply to you.

Credits and license

• Base model: deepseek-ai/DeepSeek-V4.1-Flash-1017 — MIT (its LICENSE ships here).
• GSQ-RCO quantization and the shared Engram file: ISTA-DASLab/DeepSeek-V4.1-Flash-1017-GSQ-RCO-GGUF (DASLab, IST Austria).
• Runtime: Strata.

MIT, as the base model and ISTA-DASLab's GGUFs. Built by AtomicChat.

Preservation record

Selected source bytes are verified against pinned metadata; the original eight graded mirrors also passed the previously recorded anonymous 1 MiB weight-prefix checks. This is not a full independent-device restore/load test. Hugging Face repository copies share a provider and do not guarantee permanent storage.

Collection · Source manifest · Original documentation