Runtime: Osaurus 0.25.30 or newer (osaurus.json: required_osaurus_version 0.25.30, model_version 1).

OsaurusAI/MiniMax-M3.1-Flash-JANGHT1.9

MiniMax-M3.1-Flash for 128 GB Macs. The bundle is 95.5 GiB, all of it held in memory: the model 93.3 GiB (including its 1.6 GiB vision tower) and its DSpark speculative drafter 2.1 GiB. Image input, tool calls and adaptive thinking with five reasoning efforts are supported.

A JANGHT bundle of MiniMaxAI/MiniMax-M3.1-Flash. The source is a 60-layer MoE (hidden 6144, 64 query / 4 KV heads, partial RoPE) with MiniMax sparse attention (a lightning indexer picks the top 16 blocks of 128 keys for each query), 128 routed experts per layer (top-4 plus a shared expert) on layers 3–59 and dense MLPs on layers 0–2, a 32-layer vision tower, and a 5-layer DSpark drafter. About 427B parameters outside the drafter; the routed experts are 413B (97 %) of them.

• Routed experts (84.07 GiB, 1.748 bits/weight): JANGT trellis codes, Hadamard-128 rotation with per-layer column signs, error feedback from each expert's own activation Hessian, per-row gains refit, exact outlier columns. Widths were allocated per layer and projection from each layer's measured effect on the assistant-token KL: gate/up 1 bit in 16 layers, 1.5 in 26, 2 in 15; down 1 bit in 2 layers, 1.5 in 12, 2 in 10, 2.5 in 33.
• Attention (q / k / v / o), shared experts and the dense MLPs of layers 0–2: affine 4-bit (group 64), GPTQ on each layer's captured input statistics.
• Sparse-attention indexer (index q / k projections): the release's MXFP8, repacked without requantization.
• Router gates and score-correction biases: F32 as released; norms BF16.
• Token embeddings and lm_head: affine 6-bit (group 64); the head with GPTQ on its input statistics.
• Vision tower, patch merge and projector: the release's BF16 / F32 tensors byte for byte (523 tensors).
• DSpark drafter (dspark/, 5 layers, 7-token blocks): as released (MXFP8 layers, BF16 fc and Markov head).

Bits/weight 1.878 = model bytes × 8 / source parameters (BF16-equivalent), with the vision tower included in both and the DSpark drafter excluded from both (the name rounds it to 1.9).

Quality vs the source model

Teacher-forced on held-out conversations that were never used for calibration (290 conversations, 456,954 scored positions), the bundle exactly as it ships against the source model at its release precision. KL is over the source's top-128 tokens plus one tail bucket (on assistant tokens the source's tail mass is 0.44 %).

Assistant tokens (what the model writes)

held-out domain	positions	mean KL ↓	median KL ↓	p99 KL ↓	top-1 ↑	top-5 ↑
agentic (tool calls, multi-turn)	34,066	0.128	0.022	1.58	88.8 %	99.3 %
coding	48,223	0.140	0.0008	2.41	91.3 %	99.1 %
terminal	29,650	0.048	0.0022	0.55	93.6 %	99.9 %
reasoning	29,068	0.091	0.0014	1.31	92.6 %	99.7 %
long context (documents to 16k tokens)	782	0.143	0.0072	2.17	89.5 %	99.9 %
science	16,236	0.251	0.043	2.76	84.5 %	98.3 %
story (creative writing)	36,199	0.611	0.216	5.62	71.9 %	93.0 %
multilingual	18,393	0.732	0.174	6.62	71.4 %	91.5 %
all	212,617	0.258	0.017	3.61	85.9 %	97.6 %

Every scored position (system prompts, tool schemas, tool results, user text and raw documents included): mean KL 1.236, median 0.085, top-1 70.1 %, top-5 85.4 %. That is text the model reads, not writes, and M3.1 itself is unsure there (the source's mean NLL on the held-out text is 5.44 nats per token), so small differences flip many of those positions; the KL there is also a looser bound (tail mass 12.7 %). The assistant tokens are the quality measure.

• The flips sit where the source itself is unsure. Over all positions: where the source puts ≥ 0.99 on its top token (26 % of positions) the bundle agrees 97.6 % of the time; at 0.9–0.99, 84.6 %; at 0.7–0.9, 72.6 %; at 0.5–0.7, 58.2 %; below 0.5, 45–49 %.
• Agentic, terminal, coding and reasoning hold up best. Story and multilingual writing move the most (top-1 about 71–72 %). Multilingual and science text also carried most of the trunk's error in the build's format tests.
• The 4-bit trunk was chosen by measurement. On a held-out subset (38 conversations), the 4-bit attention, shared experts and dense MLPs together cost an assistant KL of 0.015 and the 6-bit head 0.0003; the same bytes were worth more in the routed experts.
• Cybersecurity text was not in the calibration or held-out sets.

Images

• 12 / 12 graded image prompts correct: invoice OCR, shape counting, a bar chart, a table, a bug in a code screenshot, a terminal error, layout and colour, arithmetic, Chinese OCR, a two-image comparison, a 1543 × 1999 document page and a 112 px image (sizes not multiples of 28).
• Teacher-forced on the bundle's own answers with the same image features, against the source (1,764 answer tokens): mean KL 0.073, median 0.0004, top-1 93.8 %, top-5 99.7 %. Think / skip choice 12 / 12 the same as the source.
• The routed experts were calibrated with image tokens too (452 image conversations at half weight in layers 3–31). Without them an earlier build answered "Black" for a solid green image where the source answers "Green".
• One misread outside the graded fields: the invoice's shipping line $347.50 is read as $337.50.

Reasoning (adaptive)

M3.1 decides for itself, turn by turn, whether to think: the first token of every reply is <mm:think> (think) or </mm:think> (answer directly). At 327 held-out turn starts the bundle makes the same choice as the source 97.6 % of the time (the source thinks on 89.6 % of them, the bundle on 89.0 %; mean |ΔP(think)| 0.025).

domain	turn starts	same choice as the source	source thinks	bundle thinks
agentic	86	96.5 %	62.8 %	61.6 %
coding	49	91.8 %	95.9 %	95.9 %
terminal	45	97.8 %	100 %	97.8 %
reasoning	45	100 %	100 %	100 %
multilingual	40	100 %	100 %	100 %
science	33	100 %	100 %	100 %
story	25	100 %	100 %	100 %
long context	4	100 %	100 %	100 %

By reasoning effort: no effort tag 97.4 % (194 turns), low 100 % (24), medium 96.2 % (26), high 97.1 % (34), xhigh 100 % (11), max 97.4 % (38).

Raw results: evaluation/kl_vs_source_heldout.json, evaluation/reasoning_decision.json, evaluation/vision_graded.json.

DSpark speculative decoding

The bundle ships the release's DSpark drafter unchanged (5 layers reading target layers 3, 17, 31, 45 and 59; 7-token blocks; 2.1 GiB). Verification is exact: greedy output equals plain decoding, and sampled output follows the same distribution as plain sampling (the draft only proposes; the target's own sample decides every token). Greedy output with and without the drafter was token-identical on this bundle on coding, agentic, reasoning and terminal prompts. Osaurus picks the draft width each round (Adaptive).

Speed and memory (Osaurus engine)

Osaurus serving path (the app's model loading, cache and batch engine), one request, prefill step 2048, document prompts, cold cache. Medians of 6 probes per row over 2 interleaved rounds.

	M5 Max 128 GB	M5 Ultra
Prefill, 4.5k-token prompt	370-520 tok/s	1,051 tok/s
Prefill, 35.7k-token prompt	380-415 tok/s	1,134 tok/s
Decode after a 4.5k prompt	26-29 tok/s	38.3 tok/s
Decode after a 35.7k prompt	24-26 tok/s	35.9 tok/s
Peak memory, 35.7k context	99.6 GiB (102.5 with DSpark)	99.6 GiB (102.2 with DSpark)

The M5 Max was measured with the lid closed. The high end of each range is the first minutes; it settles to the low end as the chip heats. At the same GPU clock an M5 Ultra prefills about 2x faster than an M5 Max.

DSpark is on by default and is exact. On code and other predictable output it speeds up decoding: 1.24-1.29x for rewrite, list and simple-code prompts on an M5 Ultra, 1.06x on average over 15 chat prompts. On document summaries it saves nothing: 0.97x on the M5 Ultra, about 0.88x on a heat-soaked M5 Max. Turn it off in the model picker (Speculative Decoding → Off (AR)) for long-document work.

A follow-up on a cached conversation only prefills the new tokens. In the app, a follow-up question on a 35.7k-token document prefilled 188 new tokens and restored the other 36,875 from the cache. After the app unloaded the idle model and loaded it again, the next question still restored 37,059 tokens and prefilled 86.

Known limitations

• Very small images: for a 112 px solid red square (16 image tokens) with adaptive thinking, the bundle answered "Purple". The same question with thinking Off and the graded image set (which includes a 112 px image) are answered correctly.
• Nested tool arguments: asked to "run make test with CI=1 and FAST=0 and a timeout of 120 seconds" through a tool whose schema has an optional env object and an integer timeout_s, the bundle wrote CI=1 FAST=0 timeout 120 make test as the command and set timeout_s 130. The call is well-formed and valid against the schema, but the optional env object was not used; agent loops that rely on nested objects should validate the argument structure.

What's in the bundle

• Vision: the vision tower, patch merge and projector in-bundle (config.json → vision_config), with the release's image and video processors (preprocessor_config.json, image_processor.py, processing_minimax.py, video_processor.py).
• DSpark drafter: dspark/ (model-dspark-00001.safetensors + config.json); Osaurus' speculative-decoding switch turns it on or off.
• Thinking + agentic (chat_template.jinja):
  • Adaptive by default: with no thinking setting the template renders the release's adaptive instructions and adds no prefix, so the model decides each turn.
  • thinking_mode chat-template kwarg: "enabled" (always think; the reply starts after <mm:think>) or "disabled" (no thinking; the reply starts after </mm:think>), the same switch and wording as MiniMax-M3's official template. Unset, the template renders exactly as the release's. Osaurus' thinking selector maps Default / On / Off to unset / enabled / disabled.
  • Reasoning effort via the reasoning_effort (or effort) kwarg: low / medium / high / xhigh / max; no effort tag when unset.
  • Past reasoning is rendered on every earlier assistant turn, so send each turn's reasoning_content back with the history.
  • Tool calls in the release's native XML format: one <tool_call> block holding <invoke name="…"> elements with one element per argument (nested objects and arrays expand into elements), every tag prefixed by the ]<]minimax[>[ token; tool results go back in the tool role.
  • Roles: a leading root message is the high-priority system block; system / developer messages map to the lower-priority developer block. Without a root message the release's default system text is used.
• Self-describing: config.json carries the jangt expert-format block (trellis state, Hadamard rotation, outlier columns per projection) and a per-module quantization map; jang_config.json carries the format, bits / weight, source parameters, per-layer expert widths, trunk formats, embedding / head bits, sampling defaults and the drafter note. jang_config.format reads jangtq2 for loader compatibility; every routed expert is mode: jangt.
• model.safetensors.index.json covers the 23 model shards (3,416 tensors).

Serving contract

• Sampling: the release defaults, temperature=1.0, top_p=0.95
• EOS: [200020] ([e~[) · BOS: 200019 · context: 1,048,576
• Reasoning: adaptive by default; thinking_mode = enabled / disabled; reasoning_effort = low / medium / high / xhigh / max

Build details

• Source: MiniMaxAI/MiniMax-M3.1-Flash (revision 262ae68)
• Calibration: 1.86M tokens of multi-turn agentic tool use, terminal sessions, coding in 11 languages, reasoning, long-context retrieval (documents to 16k tokens), multilingual text (Aya and Wikipedia in 8 languages), encyclopedic prose, science and creative writing, rendered through the release template with think and skip turns as the model takes them and a reasoning-effort tag on about a third of the rows; the source model's own adaptive traces (424 agentic, terminal, coding, reasoning, science and multilingual prompts, the held-out ones kept for evaluation); and 452 image conversations (photos, charts, document pages, synthetic renders; 279k tokens) in the expert statistics of layers 3–31 at half weight. Held-out sets disjoint (no shared rows or prompts)
• Experts: JANGT trellis (1 / 1.5 / 2 / 2.5 bits per weight per layer and projection), Hadamard-128 rotation, Hessian-weighted error feedback, per-row gains refit, exact outlier columns; widths from measured per-layer sensitivity (one held-out run per MoE layer, assistant-token KL), allocated greedily at the byte budget
• Everything else: 4-bit GPTQ trunk, MXFP8 indexer as released, 6-bit embeddings and lm_head, each chosen by its held-out KL against what its bytes buy in the experts; vision tower and DSpark drafter as released
• License: MiniMax Community License (LICENSE), as the source

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