Runtime: Osaurus 0.25.25 or newer.

OsaurusAI/Flint-40B-A2.5B-JANG_4M

Flint 40B-A2.5B in 23.4 GB, with 96.2% top-1 agreement with the full-precision model. Against Recursive's own Apple Silicon release (RecursiveSI/flint-int4, 22.6 GB), it has 2.3× lower KL divergence from the original and 2.1 points higher top-1 agreement, at 0.8 GB more.

This is a JANG bundle of RecursiveSI/flint, Recursive's mixture-of-experts model for coding agents: 49 layers (2 dense, 47 MoE with 256 routed experts, top-6, plus an always-on shared expert), 39.8B parameters with about 2.5B active per token, 35 sliding-window and 14 full-attention layers, a 262k context, reasoning and native tool calls. Text only; the release has no MTP head or draft model.

• Routed experts: affine 4-bit (group 64), rounded with GPTQ on per-expert Hessians after activation-aware (AWQ) scaling.
• Everything else: affine 8-bit (group 64), GPTQ-rounded: attention, the shared experts, the two dense MLP layers and lm_head; the embeddings are 8-bit too. Routers, norms and the learned attention sinks stay bf16.
• Size: 23.37 GB of model shards, 4.70 bits per weight over the 39.8B source parameters.

Quality vs the original model

The test set is 96 held-out sessions (253,155 teacher-forced positions), scored against the BF16 release with top-128 KL. Every scored assistant turn was written by Flint itself; some sessions continue real agent and terminal trajectories. They cover coding in dozens of languages, multi-round agent episodes with tool calls and injected environment errors, terminal sessions, math, science, finance, security, long documents and chat. None of these sessions were used for calibration.

	size	bits/weight	mean KL ↓	p99 KL ↓	top-1 ↑	ΔNLL ↓
RecursiveSI/flint (BF16, the reference)	79.6 GB	16	0	0	100%	0
RecursiveSI/flint-int4	22.6 GB	4.54	0.0481	0.730	94.1%	+0.044
Flint-40B-A2.5B-JANG_4M	23.4 GB	4.70	0.0205	0.333	96.2%	+0.017
Flint-40B-A2.5B-JANG_6M	32.8 GB	6.60	0.0057	0.090	98.0%	+0.001
Flint-40B-A2.5B-JANGH2	13.2 GB	2.65	0.152	2.36	89.5%	+0.149

ΔNLL is the change in negative log-likelihood of the actual held-out tokens. For scale: moving every embedding row of the BF16 model by one bf16 step already gives a mean KL of 0.0040 and 98.4% top-1, so even a negligible change to the weights lands near 98% top-1.

RecursiveSI/flint-int4 is Recursive's own MLX release: affine 4-bit (group 64) on every linear layer, including attention, the shared experts and lm_head, with 8-bit embeddings (bit widths measured from its shards). It was scored on the same sessions, the same way.

By area (mean KL · top-1):

area	positions	JANG_4M	RecursiveSI/flint-int4
coding	180,831	0.0176 · 96.6%	0.0418 · 94.6%
reasoning	43,308	0.0255 · 95.4%	0.0568 · 93.2%
agentic	14,343	0.0270 · 95.4%	0.0665 · 92.7%
terminal	3,179	0.0115 · 97.2%	0.0363 · 94.8%
general, creative, long documents	11,494	0.0430 · 94.0%	0.0958 · 91.2%

Coding: review, refactor, test, port and debug tasks on real files plus generation across dozens of languages, and PyTorch. Reasoning: math, science, finance, office, networking and cybersecurity. Agentic: multi-round tool-use rollouts against a simulated environment (with errors, timeouts, empty or truncated results, permission and rate-limit failures, and mid-task user steering), next turns of real agent trajectories, and edge cases such as tools offered but not needed or prompt injection inside a tool result. Terminal: terminal-agent sessions with a native bash tool and next turns of real terminal and SWE-agent trajectories.

Reasoning and tool decisions

• Closing reasoning: at the 151 points in the held-out transcripts where Flint closed its reasoning (the first </think> of a turn), the BF16 model's top choice is </think> at 148. JANG_4M: 146. RecursiveSI/flint-int4:
• Starting a tool call: at the 74 points where Flint started a tool call (<tool_call>), the BF16 model's top choice is the call at 70. JANG_4M: 70. RecursiveSI/flint-int4: 70.

What's in the bundle

All bit widths below were measured from the shard bytes.

• Routed experts (47 MoE layers × gate/up/down × 256 experts): affine 4-bit group 64. All 36,096 expert matrices were GPTQ-rounded.
• 8-bit group 64: attention q/k/v/o (×49 each); shared-expert gate/up/down (×47 each); the dense MLP gate/up/down of layers 0–1; lm_head; embeddings.
• bf16: routers (×47), all RMSNorms including the per-head QK norms, and the attention sinks.

Serving contract

• Reasoning: Flint has no thinking switch and no effort levels. It decides each turn whether to reason and writes <think>…</think> itself before the answer; thinking toggles have no effect.
• Tool calls: the GLM-4.7 XML format, <tool_call>name<arg_key>k</arg_key><arg_value>v</arg_value></tool_call>. Tool results go back as <|observation|> turns.
• Roles: system, developer, user, assistant and observation each have their own marker; the developer role is native, not an alias for system.
• Sampling: temperature 1.0, top-p 0.95, top-k 20 (generation_config.json). Flint can reason for a long time on hard problems; allow at least 32k output tokens for them.
• Stop tokens: <|end_of_text|>, <|user|>, <|observation|>, <|developer|>.
• Context: 262,144 tokens. The 35 sliding-window layers keep only the last 513 tokens, so the KV cache grows on the 14 full-attention layers only (28 KiB per token).

Build details

• Calibration: a corpus of 548 sessions (1.80M tokens) of Flint's own outputs, generated by an 8-bit copy of the BF16 model at Flint's default sampling settings. 452 sessions (1.44M tokens) are calibration; the 96 held-out sessions above are disjoint from them. Across the corpus, 134 sessions are multi-turn, with 355 tool rounds, 134 injected environment failures and 31 mid-task user steers; coding spans 47 programming languages; and Flint chose to reason in 871 of 918 assistant turns.
• Hessians: per-expert input Hessians weighted by the router weight squared, and dense input Hessians for every other linear layer, from a layer-streamed run of the BF16 model.
• AWQ: fold strength is searched per layer on the Hessian-weighted error, with zero (no fold) allowed. The search folded the down-projection inputs through each expert's up rows in 47 of 47 layers (strength 0.2–0.4) and the expert inputs into the post-attention RMSNorm in 2, with the router and shared expert compensated exactly. The attention-input fold into input_layernorm was kept in 3 of 49 layers. Scales are powers of two, so every fold is exact in bf16.
• GPTQ: expert Hessians shrink toward the layer's pooled statistics. On held-out data this edged out each expert's own Hessian on mean KL (0.02052 vs 0.02065; top-1 96.20% vs 96.23%). A never-worse-than-round-to-nearest guard applies per expert.
• Verification: checked from the shards. Every tensor file is aligned, there is one index that matches the shards, every module's bits match the config, and no scale is NaN or infinite.