Runtime: Osaurus 0.25.21 or newer (osaurus.json: required_osaurus_version; model_version 2 = these JANGHT weights). This repository was OsaurusAI/Naive-N0.5-Flash-JANGH2 until 2026-10-13; old links redirect here. If you downloaded the JANGH2 weights earlier, this is a different bundle: 4.6 GiB smaller and closer to the original model.

OsaurusAI/Naive-N0.5-Flash-JANGHT2.5

Naive-N0.5-Flash for 128 GB Macs: 575 GiB of bf16 weights in 91.31 GiB. 4.6 GiB smaller than our JANGH2 release and closer to the original model: 29% less extra loss on the real next token.

A JANGHT bundle of NaiveAI/Naive-N0.5-Flash: a 309B MoE (15.5B active, 256 routed experts, top-8) for coding and agentic work, with a native 1M-token context built from sliding-window attention and sparse attention. Text only.

• Routed experts: JANGHT, a per-layer mix of two formats, chosen by measurement:
  • JANGT: a trellis code at 2, 2.5 or 3 bits per weight (Hadamard-128 rotation, 12-bit state, up to 8 outlier columns per layer kept in fp16). The 2.5-bit variant alternates 6- and 4-bit transitions, so a layer can sit between 2 and 3 bits.
  • JANGH: our Hadamard-32 codebook format at 2, 3 or 4 bits, rounded with GPTQ.
• Everything else: affine 8-bit, or kept in the source precision (routers in fp32, norms, the sparse-attention indexer).
• Name: JANGHT + 2.5 bits per weight averaged over the whole model (91.31 GiB × 8 / 308.9B source parameters).

Quality vs the bf16 model

124 held-out sequences (coding, cybersecurity, agentic, general, Chinese, science, academic, long documents and 96 multi-turn tool conversations), 133,881 teacher-forced positions, scored against the full bf16 model run layer by layer from disk. Top-128 renormalized KL. The official FP8 rows use NaiveAI's own deployment format (FP8 routed experts, everything else bf16, ~315 GB): the practical ceiling for any compressed version of this model.

ΔNLL is the extra loss on the token that actually comes next (nats per token, ±0.006). It is the main number for this model; agreement with bf16 has a low ceiling here (next section). Perplexity ratio vs bf16: 1.31 (JANGH2 1.46, official FP8 1.20).

Why no quantization gets close to 100% agreement here

Naive-N0.5-Flash picks 8 of 256 experts per token in every layer, and the 8th and 9th candidates are often almost tied. A change far smaller than any quantization error flips some of those choices, and the flips compound. Running the same bf16 weights in fp32 instead of bf16 arithmetic already gives only 79% top-1 agreement. That is why even NaiveAI's 8-bit deployment sits at 62% here, and why ΔNLL, which has no such floor, leads the table.

Earlier comparison: noise ceiling and a plain MLX quant (from the JANGH2 release)

68,370 teacher-forced positions on 28 held-out prompts against the bf16 model, top-128 renormalized KL. A different set from the table above (no multi-turn tool conversations), and JANGHT2.5 was not run on it.

¹ Stock MLX affine quantization of the same source (experts 2-3 bits chosen by measurement, non-experts affine 8-bit, no calibration), built as a control; not published. Served, it passed 34 / 48 tool-use cases and 11 / 20 behavior probes, against 48 / 48 and 20 / 20 for the JANGH bundles.

By domain, ΔNLL vs bf16:

JANGHT2.5 beats JANGH2 in every domain. The calibration set is weighted toward coding, tool use and cybersecurity; general, Chinese and science text lose the most.

Tool-use decisions vs the bf16 model

The 96 held-out tool conversations use tools that never appear in calibration. Decisions are scored where the model itself makes them: in its native output layout, the token after </think> and the blank line.

Live behavior (served, temperature 0)

JANGH2 passed both suites too; they check behavior and separate a working bundle from a broken one, not more.

Speed (M5 Max 128 GB, reference measurements)

Measured with the development server the bundle was validated on, one request at a time; two interleaved runs per bundle (A, B, A, B), a fresh server for each run after a 3-minute cool-down, each run the median of 3 probes on a never-seen prompt. Osaurus figures on your Mac will differ in absolute terms; the comparison between the two bundles is the point.

Decode is ~3% faster and long-prompt prefill ~15% faster than JANGH2, in a 4.6 GiB smaller bundle.

What's in the bundle

• Text only. The source model has no vision, audio or video components. No speculative drafter.

• Experts (47 MoE layers):

  projection	JANGT 2	JANGT 2.5	JANGT 3	JANGH 2	JANGH 3	JANGH 4
  gate / up	20 layers	22	—	5	—	—
  down	2	8	1	10	25	1

• Thinking + agentic:

  • Thinking is ON by default. The model opens <think> itself; with thinking off the template appends <think></think>.
  • Reasoning efforts are low / high / max (default max). The template renders any other value as Max.
  • Reasoning is kept in history by default.

• Tool calls: Qwen3-Coder-style XML (<tool_call> / <function=name> / <parameter=key>value</parameter>), declared as tool_parser: xml_function; tool results use the tool role and render as <tool_response>. Hermes-style JSON parsers will not work.

• Self-describing:

  • config.json carries the per-module quantization map: 95 JANGT expert projections ("mode": "jangt", "kbits"), 46 JANGH ("mode": "jangtq2", "bits"), 197 affine 8-bit. The jangt block holds the trellis parameters and the per-layer outlier columns; the jangtq block (version 2) holds the JANGH codebooks.
  • jang_config.json records capabilities, calibration, the allocation plan and the format of every expert unit.

• Every shard is alignment-safe (zero-copy memory mapping). 18 shards.

Serving contract

• Sampling: temperature=1.0, top_p=0.95 (vendor defaults), no repetition penalty
• EOS: 151645 · context: 1M native
• Reasoning: reasoning_effort chat-template kwarg (low / high / max), default max; declared as reasoning_parser: think_xml
• Output layout: the model writes </think>, then a blank line, then the answer or the <tool_call>.
• The routers must stay in fp32 and the sparse-attention indexer unquantized, as shipped.
• JANGH units keep their original on-disk identifiers (jangtq block, "mode": "jangtq2", *.tq2_packed / *.tq2_scales). They are not loadable by, and must not be routed to, JANGTQ v1 loaders.

Build details

• Source: NaiveAI/Naive-N0.5-Flash @ 0235b3b (bf16, 575 GiB)
• Calibration: 618k tokens rendered with the model's own chat template (coding, tool conversations, cybersecurity, agentic, general, Chinese, science, academic), referenced to the bf16 model; evaluation prompts, tools and every tenth calibration document held out
• JANGT units: column-RMS scaling and a least-squares column refit on down projections; unit-gain row scales
• JANGH units: GPTQ on the full per-expert input covariance; 31 units carried over unchanged from JANGH2, 10 down projections re-quantized
• Allocation: per projection and layer, the format and width with the lowest measured cost on held-out rows, each layer's error weighted by its measured effect on the output distribution

Quantized and validated by Jinho Jang — eric@osaurus.ai