--- license: apache-2.0 license_link: https://huggingface.co/nightmedia/Qwen3.6-27B-Akka base_model: nightmedia/Qwen3.6-27B-Akka base_model_relation: quantized library_name: mlx pipeline_tag: text-generation language: - en - zh - ja - es tags: - mlx - qwen - qwen3.5 - qwen3.6 - apple-silicon - fable - tess - speculative-decoding - quantized --- # Qwen3.6-27B — Akka, MLX 6-bit (AR-only) [Explore the model guide](https://huggingface.co/spaces/philipjohnbasile/local-ai-guide) · [All public work](https://huggingface.co/philipjohnbasile) ## Start here | Question | Answer | |---|---| | What is it for? | Studying nightmedia's Akka merge as a text-only MLX model on Apple Silicon. | | What did I contribute? | Philip John Basile's 6-bit MLX conversion, source-head calibration audit, and comparative measurements. [nightmedia](https://huggingface.co/nightmedia/Qwen3.6-27B-Akka) created the merge; its upstream lineage and credits are documented below. | | What runs it? | MLX-LM or MTPLX in autoregressive mode. No MTP sidecar is shipped and no vision tower is included. | | What was tested? | M5 Max 128 GB, macOS 27.0, MLX 0.32.0, MTPLX 2.6.0. The card records coherent AR generation, calibration failure, and same-harness perplexity comparisons. | | How much memory? | About 21.86 GB (20.36 GiB) of root tensor files. Peak memory was not measured; the runtime, OS, and KV cache need additional memory. | | What are the limits? | MTP calibration measured 0% top-1 draft agreement, so the head was omitted. The artifact is an early research derivative; broader blind behavioral qualification was not completed in the recorded audit. | ### Quick start — ordinary MLX generation ```bash mlx_lm.generate \ --model philipjohnbasile/Qwen3.6-27B-Akka-6bit-MLX \ --prompt "Explain MTP draft-head agreement testing." --max-tokens 200 ``` Use the same-harness comparison below when interpreting perplexity. Its absolute values are not interchangeable with numbers from another harness. An MLX **6-bit affine (group 64)** build of nightmedia's **Akka** merge — a Tess-based blend whose declared bases include `migtissera/Tess-4-27B`, `nbeerbower/Wichtel-Qwen3.6-27B`, `nbeerbower/CHUD-Qwen3.6-27B`, the DavidAU family (Claude-4.6-OS-INSTRUCT, Heretic2, Polar, F451), and `armand0e/Qwen3.6-27B-Fable-5-Experimental`. **Text-only.** Vision tower is not included. --- ## Why this artifact is unusual 1. **A proved negative, not an absent one.** The source carries a 15-tensor MTP draft head. We attempted full MTP contract calibration and it **failed at 0.0% draft-target agreement** — so we shipped **without** the head and we shipped the evidence of *why*. For this artifact, the failure condition is measured, documented, and reproducible: `receipts/build_outcome.json` (verdict `contract_calibration_failed`), `receipts/calibrate.json` (per-depth agreement arrays). 2. **MLX conversion of the Akka merge** — the Tess/cultural-artifact lineage made runnable on Apple Silicon with a full, honest audit trail. This is the same pipeline that produced the Fable-711 MTPLX quants, so the two generations of nightmedia/DavidAU-adjacent art are directly comparable. 3. **Comparative quality anchor, not a vibe.** WikiText-2 PPL is reported as a *pair* against the Fable-711 MLX 6-bit artifact measured in the same harness, same machine, same window budget — so the +0.139 delta means something rather than an absolute number floating alone. 4. **Reproducible to the failure.** The MTP head here was added post-merge via script (per its author's own notes) and was never trained on this body. The receipts record the exact agreement numbers, which establishes the calibration failure. Training a compatible head is a possible follow-up, not a demonstrated fix. --- ## Scope — read before assuming anything - **AR-only. No MTP. No speculative acceleration.** MTP contract calibration against this merged trunk failed at **0.0% top-1 draft agreement** (0/64 at every tested depth; even top-8 only 2/64; `best_agreement = 0.0`). Shipping that head would produce a model that "loads and generates fine" with a silently non-functional draft head — the worst failure mode, because nothing errors. It is therefore **not included**. Do not expect MTP or MTP-style speedups. - **No vision.** The source is `image-text-to-text` (333 vision tensors). This conversion is the language body only; `vision_config` is absent. - **Not a quality claim.** This is an early derivative intended as a health checkpoint and study artifact; a fresh blind behavioral eval is not yet complete (refusal/scoring harness runs planned). --- ## What is included - 21.86 GB (20.36 GiB), 5 safetensors shards, **6-bit affine, group 64**. - The language body is complete: **all 848 source body-layer tensors are present; 0 dropped** (1847 tensors total in the artifact index, including affine scale/bias companions). - Verified on-device before upload: full index/shards resolve 1847/1847, AR generation coherent on the M5 Max. --- ## Usage Plain MLX checkpoint — no MTP sidecar, no special flags: ```bash mlx_lm.generate --model philipjohnbasile/Qwen3.6-27B-Akka-6bit-MLX \ --prompt "Explain MTP draft-head agreement testing." --max-tokens 200 ``` Serve AR-only through MTPLX (auto falls to AR when no MTP contract is present): ```bash mtplx serve --model philipjohnbasile/Qwen3.6-27B-Akka-6bit-MLX --generation-mode ar ``` --- ## Measured — M5 Max (128 GB), macOS 27.0, mlx 0.32.0, mtplx 2.6.0 | Check | Result | |---|---| | AR decode (`mtplx serve`, 256 tok, non-streaming) | **20.1 tok/s** (client end-to-end ≈ 20.1) | | WikiText-2 PPL, ctx 2048, 120 windows (this repo's harness) | **6.386** | | Same-harness delimiter: Fable-711 MLX 6-bit | **6.247** (Δ **+0.139**, ≈ +2.2%) | | MTP agreement (source head vs trunk), depths 1–3 | **0.0%** (top-1 0/64; top-8 2/64) | The PPL harness here (`bench-lounge/ppl.py`) is a *new* implementation and is **not** numerically comparable to the 5.7358 figure on the Fable-711 card. Read the **delta** between the two same-harness rows, not the absolutes. No peak-memory figure is claimed — it was not measured on this build. --- ## Receipts (in-repo, `receipts/`) | File | Contents | |---|---| | `build_outcome.json` | Forge verdict `contract_calibration_failed`; full probe diagnostic | | `calibrate.json` | Per-depth agreement arrays (top-1/2/4/8 rates, 0.0 shown) | | `ppl_akka6.json` / `ppl_fable6.json` | Same-harness PPL pair | | `serve_ar.log` | The `mtplx_openai_generation` line (19.9–20.1 tok/s) | Local build receipts (not uploaded): staged manifest SHA-256 for `model.safetensors.index.json` = `d64de023…`, `config.json` = `63225a56…`. --- ## Reproducibility - Source pinned: `nightmedia/Qwen3.6-27B-Akka` (bf16, 12 shards). - Forge recipe: `body_bits=6`, `body_group_size=64`, `body_mode=affine`, `body_dtype=bf16`, MTP contract calibration **enabled and allowed to fail** (it did). Same forge pipeline as the Fable-711 MTPLX 4/6/8-bit quants. - Sequence: `mtplx forge build --repo …/Akka-src --recipe <6bit> --branded-name Qwen3.6-27B-Akka-MTPLX-6bit` → `contract_calibration_failed` → AR-only package published here. --- ## Credits - **nightmedia** — the Akka merge and its source card (`nightmedia/Qwen3.6-27B-Akka`). Tess lineage anchor: `migtissera/Tess-4-27B`. - **Qwen** — Qwen3.6-27B base. **MTPLX** — forge/quant/Serve runtime. - Package and audit by philipjohnbasile; see the linked discussion on the source thread for the calibration methodology. Apache-2.0, inherited from the source model. --- ## Intended use / limits - Research, study, and creative work — not a product claim. - No MTP speedup; treat AR-only throughput (~20 tok/s on M5 Max class) as the ceiling for this artifact. - If you need a *working* MTP build of this lineage, the head requires fine-tuning onto the merged body first — the calibration failure is the definition of that requirement, and it is now on record.