license: other
license_name: liquid-foundation-model-community-license
license_link: https://www.liquid.ai/community-license
base_model: DuoNeural/LFM2.5-8B-A1B-Hermes-Agentic-Coder-Abliterated-v3
tags:

• duo-neural
• agentic
• coding
• function-calling
• hermes
• liquid-foundation-model
• mlx
  language:
• en
  library_name: mlx

DuoNeural v3 — MLX (4-bit)

MLX conversion of DuoNeural/LFM2.5-8B-A1B-Hermes-Agentic-Coder-Abliterated-v3, an 8.3B-total / 1.5B-active-parameter LFM2.5 MoE agentic-coding model. This is the 4-bit build — the smallest and fastest, at 5.0 GB it runs comfortably on any 16 GB Apple-silicon Mac:

Usage

pip install mlx-lm
mlx_lm.generate --model phntmwvs/DuoNeural-v3-4bit-MLX \
  --prompt "Write a Python function that reverses a string."

Or serve an OpenAI-compatible endpoint:

mlx_lm.server --model phntmwvs/DuoNeural-v3-4bit-MLX --port 8080

Conversion

• Source: DuoNeural/LFM2.5-8B-A1B-Hermes-Agentic-Coder-Abliterated-v3 @ d148c378cc1afc2e335f6b94a1e83f322a1b6fb0
• Tool: mlx-lm 0.31.3_3 on Apple M4 Pro (48 GB), macOS 27.2
• Quantization: affine 4-bit, group 32; embeddings bumped to 6-bit/g64; MoE router gates held at 8-bit/g64 (per mlx-lm's lfm2_moe quant predicate)
• Full provenance in CONVERSION.md; load/generate/benchmark evidence in verify.log.

Performance (M4 Pro 48 GB)

• Generation: 160.9 tok/s · prompt: 146.2 tok/s · peak memory: 5.4 GB

Evaluation

Evaluated on a 3-component harness (phntmwvs/duoneural-v3-eval-harness, all runs on the same M4 Pro):

EvalPlus detail:

Notes on reading these numbers: BFCL multi-turn is all-or-nothing per conversation (200 convs/category); ±1 pp is run-to-run noise. The Hermes FC suite is 40 hand-authored cases (20 single / 10 parallel / 5 negative / 5 system2), scored by exact tool-name + recursive argument match; this 4-bit row's +10 pp over BF16 is within small-suite variance, and we read the ladder as "quantization costs nothing measurable" rather than "4-bit is better." The real cost shows on EvalPlus (−1.9 pp vs BF16) — still above DuoNeural's published figures on every metric.

Cross-check vs vendor claims: local EvalPlus rows exceed DuoNeural's published figures on all four metrics (e.g. HumanEval+ 53.05 vs 50.0 published for this 4-bit build) — published numbers treated as conservative; vendor methodology differs (our runs use --greedy, n=1, temp=0, chat mode). A stock-base BFCL cross-check was attempted and dropped: the stock LiquidAI base speaks a different function-calling dialect (<|tool_call_start|> family) than this v3 fine-tune (Qwen/Hermes <tool_call>), so a single-handler comparison measures template mismatch, not capability.

Verified before release

• Load test + generation samples (instruction-following, strict-JSON tool call, code benchmark) — verify.log
• Tokenizer/chat-template shipped unchanged from source (chat_template.jinja, tokenizer_config.json)
• All three variants ran the full 3-component eval matrix above

Thanks

A massive thanks to Aura, Archon, and Jesse at DuoNeural Research Lab for providing the base model :) I learned a lot doing this (and still have very much more to learn!)

License

LFM Open License v1.0 (asserted via the source model card; no LICENSE file shipped in the source repo). Apache-style grant; commercial use permitted only below $10M annual revenue — at/above that threshold a separate agreement with Liquid AI is required. Redistribution requires retaining notices.