Qwen3.8 27B Cyber

Q6_K · Q8 MTP + projector · Measured configuration

66.98 / 100 Mango weighted score · 45.60 min matched workload · 12/12 mechanical agent passes

philbert440's Q6_K cyber-specialist GGUF, preserved together with its separate Q8 MTP head and Q8 vision projector.

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

The Cyber v2 Q6+Q8 MTP configuration scored 66.98/100 in 45.60 workload minutes, with zero infrastructure errors or length finishes. It passed local coding 17/20, function calls 27/40, instructions 14/20 and 12/12 mechanical cases.

Social/EQ 38.68 combines pure EQ 32.91 and OpenToM 8/18. Three of 30 EQ outputs failed parsing/range checks and five formula values were negative; the upstream formula permits this and no clipping was applied. Publisher BF16 results do not replace these exact-configuration observations.

It trails general Qwen MLX4 by 13.66 points, with a paired 95% interval [-18.91,-8.23], and previous Flash by 17.21[-22.72,-11.85]. This sample supplies no broad quality/speed case for promotion. Its different published lineage and separate draft also prevent reading the IQ4/Q6 gap as a controlled quantization result.

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 numerical report contains aggregate statistics and hashes; no prompts or generated answers are published.

Runtime and compatibility

The main model, optional MTP head and vision projector have separate roles. Loading a text model does not validate image input or speculative decoding. Each path needs a compatible runtime and its own preflight.

202 matched questions plus 12 mechanical cases; initial 32768 configured context, q8_0 KV cache, native medium (API high mapping), separate Q8 MTP draft length 3. Native binary on restored host: 0.7.0-dev, build 1, commit 63b4b303b. Context headers absent; native preflight/profile provide configuration evidence.

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

No separate companion mirror is declared in this selected manifest. This does not imply that every optional upstream projector, draft head or multimodal path is included; consult the exact file list.

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-qwen27-cyber-q6k \
  Qwen3.8-27B-Uncensored-Cyber-Q6_K.gguf \
  mtp-Qwen3.8-27B-Uncensored-Cyber-Q8_0.gguf \
  mmproj-Qwen3.8-27B-Uncensored-Cyber-Q8_0.gguf \
  --revision 7ebd93f22390961170a4fdf6b642f133a081b30c \
  --local-dir ./weights/qwen27-cyber-q6k

Load example and setup scope

llama-server \
  --model ./weights/qwen27-cyber-q6k/Qwen3.8-27B-Uncensored-Cyber-Q6_K.gguf \
  --host 127.0.0.1 --port 8080 --ctx-size 32768 \
  --jinja --flash-attn on --cache-type-k q8_0 --cache-type-v q8_0

This is a minimal text-serving example for a compatible build, not a replay of every benchmark fitting/offload flag. The recorded configuration and engine caveats define the tested scope.

Provenance and license

The source calls this the Cyber v2 line and reports BF16, Claude-judged results. Those are publisher-side results on a different precision and protocol; they are not a Mango Q6_K benchmark.

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.

Qwen3.8-27B-Uncensored-Cyber — GGUF

llama.cpp GGUF quants of Qwen3.8-27B-Uncensored-Cyber (v2 recipe), with the vision projector (mmproj-*) for image input and the MTP head (mtp-*) for speculative decoding.

Files

Full BF16 is not shipped as a single GGUF (exceeds the 50GB per-file limit) — use Q8_0 (≈lossless) or the bf16 safetensors repo.

Evaluation (bf16, Claude-judged; cyber = 100 held-out cyber-offensive prompts, regex refusal harness)

Note

Uncensored / de-refused, tuned to fully answer cyber and offensive-security questions. Use responsibly and in compliance with applicable law.

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