Instructions to use philipjohnbasile/Qwen3.6-27B-Akka-6bit-MLX with libraries, inference providers, notebooks, and local apps. Follow these links to get started.

Libraries

How to use philipjohnbasile/Qwen3.6-27B-Akka-6bit-MLX with MLX:

# Make sure mlx-lm is installed
# pip install --upgrade mlx-lm

# Generate text with mlx-lm
from mlx_lm import load, generate

model, tokenizer = load("philipjohnbasile/Qwen3.6-27B-Akka-6bit-MLX")

prompt = "Write a story about Einstein"
messages = [{"role": "user", "content": prompt}]
prompt = tokenizer.apply_chat_template(
    messages, add_generation_prompt=True
)

text = generate(model, tokenizer, prompt=prompt, verbose=True)

Notebooks
Google Colab
Kaggle
Local Apps Settings
LM Studio

How to use philipjohnbasile/Qwen3.6-27B-Akka-6bit-MLX with Pi:

Start the MLX server

# Install MLX LM:
uv tool install mlx-lm
# Start a local OpenAI-compatible server:
mlx_lm.server --model "philipjohnbasile/Qwen3.6-27B-Akka-6bit-MLX"

Configure the model in Pi

# Install Pi:
npm install -g @earendil-works/pi-coding-agent
# Add to ~/.pi/agent/models.json:
{
  "providers": {
    "mlx-lm": {
      "baseUrl": "http://localhost:8080/v1",
      "api": "openai-completions",
      "apiKey": "none",
      "models": [
        {
          "id": "philipjohnbasile/Qwen3.6-27B-Akka-6bit-MLX"
        }
      ]
    }
  }
}

Run Pi

# Start Pi in your project directory:
pi

MLX LM

How to use philipjohnbasile/Qwen3.6-27B-Akka-6bit-MLX with MLX LM:

Generate or start a chat session

# Install MLX LM
uv tool install mlx-lm
# Interactive chat REPL
mlx_lm.chat --model "philipjohnbasile/Qwen3.6-27B-Akka-6bit-MLX"

Run an OpenAI-compatible server

# Install MLX LM
uv tool install mlx-lm
# Start the server
mlx_lm.server --model "philipjohnbasile/Qwen3.6-27B-Akka-6bit-MLX"
# Calling the OpenAI-compatible server with curl
curl -X POST "http://localhost:8000/v1/chat/completions" \
   -H "Content-Type: application/json" \
   --data '{
     "model": "philipjohnbasile/Qwen3.6-27B-Akka-6bit-MLX",
     "messages": [
       {"role": "user", "content": "Hello"}
     ]
   }'

Hermes Agent

How to use philipjohnbasile/Qwen3.6-27B-Akka-6bit-MLX with Hermes Agent:

Start the MLX server

# Install MLX LM:
uv tool install mlx-lm
# Start a local OpenAI-compatible server:
mlx_lm.server --model "philipjohnbasile/Qwen3.6-27B-Akka-6bit-MLX"

Configure Hermes

# Install Hermes:
curl -fsSL https://hermes-agent.nousresearch.com/install.sh | bash
hermes setup
# Point Hermes at the local server:
hermes config set model.provider custom
hermes config set model.base_url http://127.0.0.1:8080/v1
hermes config set model.default philipjohnbasile/Qwen3.6-27B-Akka-6bit-MLX

Run Hermes

hermes

Atomic Chat

OpenClaw

How to use philipjohnbasile/Qwen3.6-27B-Akka-6bit-MLX with OpenClaw:

Start the MLX server

# Install MLX LM:
uv tool install mlx-lm
# Start a local OpenAI-compatible server:
mlx_lm.server --model "philipjohnbasile/Qwen3.6-27B-Akka-6bit-MLX"

Configure OpenClaw

# Install OpenClaw:
npm install -g openclaw@latest
# Register the local server and set it as the default model:
openclaw onboard --non-interactive --mode local \
  --auth-choice custom-api-key \
  --custom-base-url http://127.0.0.1:8080/v1 \
  --custom-model-id "philipjohnbasile/Qwen3.6-27B-Akka-6bit-MLX" \
  --custom-provider-id mlx-lm \
  --custom-compatibility openai \
  --custom-text-input \
  --accept-risk \
  --skip-health

Run OpenClaw

openclaw agent --local --agent main --message "Hello from Hugging Face"

Qwen3.6-27B-Akka-6bit-MLX / README.md

philipjohnbasile

Clarify release scope and connect the model portfolio

1bad3b5 verified 9 days ago

preview code

Raw

History Blame Contribute Delete

8.09 kB

	---
	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.