NInfer-all

One line of NInfer for the RTX 3090, RTX 4090, RTX 5090 and RTX PRO 6000 Blackwell: the forks that carry it, consolidated into one tree, plus this repository's own work. Every change keeps its author; the maintainer map lists them with the files they touch.

Where the code comes from

The base is the master of ashalliants/ninfer-3090: v0.12.0 (prompt grafts, /slots session persistence, the effective thinking budget, worker recovery) and the multi-GPU pipeline stages, most of both by Warlax, on the line Don-Chad/ninfer-3090 started from Neroued's NInfer. On top of it come:

• patches from TertiumOrganum1/ninfer-3090;
• ideas from UDPSendToFailed/ninfer-4090 and its contributors;
• open pull requests to Neroued/ninfer;
• work by IMGillusion, Mirko Covizzi, Ian Ranson (Wallawalla47), tmark00 and David Oelfke (gzenz/ninfer).

What the engine does beyond this page (packages, serving APIs, supported models, flags) is in the READMEs of NInfer-3090 and NInfer-4090. New features that change numbers or serving behaviour are opt-in, except four defaults:

• the routes the card's measured device profile picks (--device-profile off keeps the compiled tables);
• prefill chunks rounded to whole waves of the card's SMs (NINFER_PREFILL_ALIGN=0 keeps the requested chunk);
• TertiumOrganum1's ternary prefill tile (NINFER_T2_A8_TILE=off);
• endpoint anchors (--no-endpoint-anchors).

Highlights

Ternary Bonsai 2 27B and Qwen3.8-27B

Measured in September 2026, one card each, greedy, one request unless the row says otherwise. Setups and full tables: reference measurements.

	RTX 3090	RTX 4090	RTX 5090	RTX PRO 6000
Ternary Bonsai 2 27B, short chat (DFlash2, 7 drafts)	202 tok/s	256 tok/s	397 tok/s	381 tok/s
decode after a 261K-token document (fastest drafter)	90 tok/s	123 tok/s	218 tok/s	218 tok/s
time to first token for a 261K-token prompt	215 s	102 s	82 s	78 s
largest context, filled and all three needles found	970,752	958,464	978,944	1,048,576*
eight requests at once (MTP, 3 drafts), total	551 tok/s	824 tok/s	1,063 tok/s	1,155 tok/s
Qwen3.8-27B, short chat (DFlash2, 7 drafts)	118 tok/s	149 tok/s	236 tok/s	237 tok/s
largest context, filled and all three needles found	417,792	405,504	872,448	1,048,576*
eight requests at once (MTP, 3 drafts), total	329 tok/s	442 tok/s	690 tok/s	739 tok/s

* The engine's ceiling. The RTX PRO 6000 (96 GB) starts there with every KV format and drafter; filled to it, both models find two of the three needles.

• RTX PRO 6000 against RTX 5090. Re-measured in October beside an RTX 5090, both at 600 W, the PRO 6000 came back within 2% of September wherever the drafts accepted the same share. One request's decode is bound by memory bandwidth, and both cards have 1.79 TB/s of GDDR7 (Bonsai 2 without speculation: 167.7 against 170.4 tok/s). The PRO 6000's 188 SMs against 170 show where compute decides: the 261K prompt is 2% faster and eight requests at once 6 to 7% faster (re-check).
• Against the previous master on the same card. A 261K-token Bonsai prompt takes 215 s instead of 315 s on the RTX 3090, 102 s instead of 138 s on the RTX 4090 and 82 s instead of 115 s on the RTX 5090; rk4v4 decode after it is 11 to 13% faster on the 24 GB cards. Decode at short context is unchanged, and Qwen3.8's 8K to 32K prompts on the RTX 5090 take 8 to 10% longer.
• Draft length. DFlash2 with seven drafts is fastest on short answers; after long documents the best count is three to seven. MTP runs up to fifteen drafts and is fastest at three to five.
• Past the native window. Filled to about 880K tokens, Bonsai 2 returned all three planted codes on every card. At 1,048,576 tokens (RTX 5090 and PRO 6000 only) it misses the one at 943K.

Qwen3.8-Flash-Next

One request, greedy, October 2026. Each cell is decode of a short answer · prefill of a 4,463-token prompt, in tok/s. Decode after the long prompt, memory, host links and power limits: Qwen3.8-Flash-Next.

Q2_0 (35.9 GiB model + 26.8 GiB n-gram table):

Experts	RTX PRO 6000	RTX 5090	RTX 4090	RTX 3090
on the GPU (two cards, except the PRO 6000)	138 · 2,939	136 · 3,809	97 · 3,608	90 · 1,504 †
in pinned host memory	55 · 1,297	74 · 1,677	53 · 1,138	49 · 842
on disk, the files in the page cache	76 · 2,024	68 · 1,739	42 · 723	47 · 630
on disk, cold (pages evicted every second)	35 · 1,000	25 · 694	18 · 160	17 · 195

† Two RTX 3090 Ti.

IQ3_S (51.9 GiB model + the same table):

Experts	RTX PRO 6000	RTX 5090	RTX 3090
on the GPU (two cards for the RTX 5090)	125 · 2,541	122 · 3,326	—
in pinned host memory	29 · 803	41 · 1,100	34 · 533
on disk, the files in the page cache	66 · 1,661	55 · 439	19 · 209 ‡
on disk, cold	29 · 773	19 · 171	11 · 47

— Not measured: every expert needs three 24 GB cards. ‡ The host's 62 GB of RAM cached only part of the file.

Coder IQ1_M (28.4 GiB model + the same table), one NVIDIA L40S (48 GB):

Experts	NVIDIA L40S
in pinned host memory	35 · 1,648
on disk, the files in the page cache	42 · 1,938

• The host matters. Host and disk rows depend on the host as much as on the card. The RTX 5090 host had PCIe 5.0 x16; the others had PCIe 4.0 x16.
• The PRO 6000 caches almost everything. Its 96 GB device expert cache ends up holding nearly every expert.
• RTX 4090 host pinning. The RTX 4090 rows ran pinned to the GPUs' NUMA node in a two-socket VM. Unpinned, host experts decode there at 40 tok/s and disk experts at 32 to 33.
• Many requests at once. Six reasoning requests at once on two RTX 5090s produced 193 tok/s of output with Q2_0.

What this line adds

• Model suspend. --model-suspend lets an idle server give its device memory back without exiting (POST /v1/models/{id}/suspend) and take it again on the next request or on /resume, retained conversations included. Ternary Bonsai 2 27B on an RTX 3090: 7.9 GiB down to 0.3 GiB in 0.33 s, back in 1.1 s. Output and prefix reuse are identical afterwards, on one device or across pipeline stages. Model suspend.

• Several models behind one server. With --models-dir or a llama.cpp-style --models-preset, ninfer-serve is a router with llama.cpp's model API (/models, /models/load, /models/unload, /models/sse, --models-max). With --model-suspend a model sleeps instead of unloading: two 27B models on one 24 GB RTX 3090 swap in 1.8 s instead of an 11.6 s cold load. Several models.

• llama.cpp's native endpoints. POST /completion and /v1/completions continue a raw prompt (text, token ids or both). /tokenize, /detokenize and /apply-template expose the tokenizer and the chat template. Raw-prompt completion.

• Rerank. POST /v1/rerank (Jina, llama.cpp and TEI shapes) ranks documents with the served model as the judge, scored Qwen3-Reranker's way: P(yes) / (P(yes) + P(no)). Rerank.

• GGUF block formats. Qwen3.8-27B GGUF releases that pick a ggml type per tensor, such as ISTA-DASLab's GSQ-RCO, convert without requantization (qwen3_8_27b_gguf), with MTP, DFlash2 and Vision. The 3.5-bit IQ3_S release:

  • 10.95 GiB of weights instead of 15.9;
  • WikiText-2 perplexity 7.071 (its card: 7.07; the official artifact: 7.286);
  • 80.3% on IFBench, 100% on AIME 2025 and 2026, 88.4% on GPQA-Diamond (the official artifact: 77.7, 96.7, 96.7 and 87.4);
  • 59.9 tok/s against the official artifact's 40.3 on an RTX 3090.

  GGUF block formats.

• Device route profiles. Each card's measured profile picks the kernel route for every operation and width before the compiled tables do. Profiles are built in for the RTX 3090, 4090, 5090 and the three RTX PRO 6000 editions; any other GPU is measured once at first start (20 to 40 s), and ninfer-calibrate re-measures. On the RTX 3090 the rk4v4 verify attention at 262K runs 3.2 times faster, and FP16 P·V with the fast prompt kernel cuts prompt attention by 19 to 30%. A greedy answer served in a batch now parts more often from the same request served alone, at near-tied tokens. --device-profile off keeps the compiled routes. Device profiles.

• FP8 and NVFP4 on the default Blackwell build. Every 120a build carries the FP8 A8 and NVFP4 W4A4 tensor-core units. On an RTX PRO 6000, Qwen3.8-27B NVFP4/FP8 prefills 4,096 tokens at 11,822 tok/s, within 1.4% of a native build; Qwen3.6-35B-A3B NVFP4 prefills at 30,938 tok/s.

• Faster attention at long context. Three changes cut a 131K rk8v4 prompt on an RTX 3090 from 101 s to 76 s:

  • new small-T tiers for the INT8-family caches;
  • the fast prompt kernel for rk8v4, rk4v4, rk4v4-e8 and rk2v4-e8;
  • prefill chunks sized to whole SM waves.

• MTP up to fifteen drafts. --draft-tokens 10..15 starts; it used to fail at graph update.

• BF16 KV with graphs. MTP with the default BF16 KV cache, and Qwen3.8 without speculation at 512 and 1,024 tokens of context, no longer fail at startup.

• Parallel query tiles (opt-in). A verify step or a short prefill over an INT8-family cache runs its 9 to 64 columns as tiles of one split-KV launch (attn_parallel_tiles in the profile, or NINFER_ATTN_PARALLEL_TILES=1).

• Branch and endpoint anchors.

  • --branch-anchors (opt-in) captures a request where its prompt stops matching a retained conversation.
  • Endpoint anchors (on by default; --no-endpoint-anchors) keep the point a continued turn resumed from. Another reply to the same answer, or an edited last message, resumes there. On an RTX 4090, such a branch of a 30.7K-token conversation answers in 110 ms instead of 9.5 s, for 7-8 ms more on the continued turn.

• Blackwell kernels (120a builds only):

  • MX FP8 MMA with TMA split-K for FP8 A8 projections;
  • FP4 tensor-core QK for an NVFP4 KV cache past 2,048 keys (--fast-prefill-kernel);
  • programmatic dependent launches in captured graphs;
  • native FP8 and NVFP4 A16 operands with CUDA 13.2.

• Qwen3.8-Flash-Next. ISTA-DASLab's GSQ-RCO GGUF releases of the 125B-parameter MoE (512 experts, about 6B active) convert without requantization. The experts can live:

  • on the GPU, or on the GPUs of a --devices pipeline;
  • in pinned host memory with a GPU cache (--expert-residency host);
  • in the file, read into a GPU cache (--expert-residency disk, under 1 GB of RAM).

  It serves up to eight requests with prefix reuse, structured output, images and video; no release carries an MTP layer, so there is no MTP. On two RTX 5090s the Q2_0 release scores 93.3% on AIME 2025 and 86.4% on GPQA-Diamond (84.3% within 106,000 output tokens; its card, from llama.cpp: 96.67 and 89.39). Qwen3.8-Flash-Next.

• Ternary Bonsai 2 27B. PrismML's ternary Qwen3.8-27B runs from t2_g128_fp16 weights: 2.125 bits per weight, imported without rounding, with the Hadamard rotations fused into the norms and gates. The recipe bonsai2_27b_ternary adds ProCreations' MTP head and DFlash2 adapter and an exact proposal head.

• Integer activations for ternary projections. Decode, verification and prompts up to 192 tokens use a small-T kernel over s8 activations; longer prompts use the int8-activation GEMM.

• RTX 3090 tuning. Four warps per 1024-point rotation, small-T attention splits in whole SM waves, the GDN record window in shared memory, the ternary target's DFlash2 adapter in Q4.

• DFlash2 with Vision in overlay. An image encode borrows the drafter's memory, so DFlash2, Vision and the full 262,144-token window fit on one 24 GB card.

• Serving fixes.

  • A forced tool_choice opens the named call.
  • A context-cache store that cannot place a request fails only that request (HTTP 429).
  • A Paged KV exhaustion names its page numbers, and three in a row mark the engine unhealthy.
  • Context-cache fixes keep long agent sessions from re-prefilling.

• Build. Tests build against CUDA 13's cudaGraphGetEdges; compressed device code keeps the binaries under 2 GiB.

• Reference measurements of Bonsai 2 and Qwen3.8-27B on four cards: September 2026.

From other forks

From TertiumOrganum1's fork: the rk4v4-e8 KV cache, the ternary prefill tile, tool-call recovery

From TertiumOrganum1/ninfer-3090:

• rk4v4-e8 KV cache. Keys rotated as in rk8v4 and snapped per octet to the E8 lattice in int4 (the E8 codecs first appeared in NInfer-4090, by UDPSendToFailed with Daniel Parker); values keep rk8v4's int4 plane. 280 bytes per token and KV head instead of 408. On Ternary Bonsai 2 the 262,144-token window takes 2.0 GiB less and two lanes get a whole window each; the codes planted at 131K and 250K are still found, and quick-corpus perplexity moves from 5.631 to 5.650.
• A 128x64 int8 tile for ternary prefill. Activations are quantised per token and 128-column group. Bonsai 2 prefill runs 33% faster at 8K, 21% at 32K and 15% at 64K, perplexity unchanged (5.631). NINFER_T2_A8_TILE=off restores the old kernel.
• Tool calls. A malformed tool-call region is recovered as far as it reads instead of leaking its markup into the answer.
• Shared captures. A capture that releases less than was assessed is abandoned; before, the engine failed for good and answered 503 until a restart.
• Build. sm_120a builds on the mma.sync compatibility path.

From NInfer-4090: whole-program build, keys past 262,144 with YaRN, the rk2v4-e8 KV cache

From UDPSendToFailed/ninfer-4090 (UDPSendToFailed unless named), re-implemented here:

• Whole-program CUDA build (Matt Anderson). No relocatable device code in the core and ops archives: a fifth fewer kernels need a stack frame; the server binary grows by a quarter.
• Shared-memory scale reads in the INT8-family attention kernels. With the whole-program build the rk8v4 verify attention takes 9 to 21% less time: on Bonsai 2 an MTP step at 64K is 6.6% shorter and prefill 2 to 7% faster from 8K up, with the same answers.
• TCP_NODELAY on the server socket.
• Sigmoid, SiLU and softplus on the SFU (opt-in). -DNINFER_SFU_SIGMOID_SILU=ON: Bonsai 2 prefill 2% faster from 8K up, perplexity 5.6306 → 5.6309, MTP decode unchanged. -DNINFER_SFU_SOFTPLUS=ON does the GDN decay gate the same way, with a log1p series for slow decays (5.6302).
• Keys past 262,144. The small-T attention kernels read page indices from the block table past the 64 they stage; the visible-key limit is 1,048,576.
• Four times the native window. --max-context up to 1,048,576 on the 262,144-token models. Past the native window positions run plain RoPE, or YaRN with --rope-yarn at Qwen's factor (--rope-yarn-factor F fixes it). On Bonsai 2 with rk2v4-e8, plain RoPE found all three codes at 500,000 tokens; YaRN found two of three at 131,072, 500,000 and 1,000,000, so it stays off unless plain RoPE stops answering.
• rk2v4-e8 KV cache (with Daniel Parker, Neroued#173). Two bytes per 8-dimension key block: 216 bytes per token and KV head, the one format that holds 1,048,576 tokens beside Bonsai 2 on a 24 GB card. The cost: perplexity 5.631 → 5.820 (rk4v4-e8: 5.651), DFlash2 acceptance 54.4% → 51.8%, decode 4% slower. All three codes are found at 131K, 250K and 500K.
• D3D12-resident arenas on Windows (with keylimesoda). -DNINFER_D3D12_RESIDENCY=ON adds --wddm-evictable-budget. Untested: this line has no Windows machine, and the code only passes a MinGW syntax check.
• Also: a server default reasoning effort, MTP draft windows up to 15, /metrics, /slots (Sergiusz Michalik) and /props, a WebUI compiled in from NINFER_WEBUI_DIR, the block sampler's candidates in shared memory, an opt-in bf16 residual add (-DNINFER_BF16_RESIDUAL_ADD=ON), vector stores in the chunked GDN prefill, and bounded split compilation with ptxas reports.

From other forks and upstream pull requests: disk KV tier, adaptive MTP, structured output, log probabilities, n-gram drafting, unified Linear templates and more

• Disk KV tier (IMGillusion). --disk-kv-path DIR writes evicted conversations' KV pages and state to CRC-checked LRU files that survive restarts; --disk-kv-restore seeds a request from a stored prefix. On Bonsai 2 with MTP a 17,444-token prompt comes back in 1.2 s instead of 9.6 s (1.0 s after a restart). On Windows, -DNINFER_DIRECTSTORAGE=ON reads restores through DirectStorage (untested).
• Adaptive MTP (Mirko Covizzi). --adaptive-mtp verifies 3..K of the K drafts per round, by measured draft survival and round cost. On an RTX 3090 with K=5 it did not beat a fixed K=3 (200 against 204 tok/s on short prompts, 150 against 163 at 8K), and its per-width graphs cost memory.
• Fast INT8 prompt attention (Ian Ranson, Wallawalla47). Registers hold each warp's rows, scores and output; P·V accumulates in FP16 per 64-key tile. This line extends it to rk8v4 and the packed key codings, and the device profiles turn it on (19 to 30% less prompt-attention time); --fast-prefill-kernel forces it. Perplexity at 64K moves from 5.2074 to 5.2079. An nvfp4 KV cache on Blackwell has its own fast kernel with QK on FP4 tensor cores: 3.5 to 14.4% faster prefill at 16K-64K on an RTX 5090.
• Agent-harness tool calls. <function name=...>, <invoke name=...>, <function_calls> and <param name=...> forms (upstream PR #300 by Pavel Kochubey, via Wallawalla47).
• Structured output through xgrammar, speculation included: --structured-output (upstream PR #294 by Andrey Shvartsman).
• Token log probabilities. logprobs/top_logprobs in Chat Completions and message.output_text.logprobs in Responses, up to 20 alternatives, streamed or not (Fedor Suchkov's frinfer design).
• Cache policies. --context-cache-policy rolling rolls one long conversation's frontier forward (IMGillusion); --release-diverged-checkpoints drops first the checkpoints a conversation has moved away from (Ian Ranson, after pkochubey's upstream PR #300).
• NVFP4 expert banks (upstream PRs #286-#290 by Mykhailo Dementii). Qwen3.6-35B-A3B NVFP4 converts with --recipe qwen3_6_35b_a3b_nvfp4; an RTX 5090 (native build) prefilled 27,663 tok/s at 4K and decoded 397 tok/s. W4A4 prefill exists only on Blackwell, so sm_8x builds refuse the banks.
• N-gram copy drafting (remesis, Ian Ranson). A round may verify up to 15 tokens copied from earlier text that matches the last 12; on by default with --spec, exact since the target verifies every copy (--ngram-draft-tokens, --ngram-min-match, --ngram-archive-mib).
• Hybrid prefix cache (Ian Ranson). --use-alt-prefix-caching: content-addressed 64-token KV blocks plus sparse state snapshots. Around the default catalog: --recency-eviction, --kv-lease-growth, --host-cache-mib, --auto-long-anchors; on by default, reuse of an aborted request's prefill and least-recently-used replacement of automatic shared prefixes.
• Admission and eviction (Gideon Zenz, David Oelfke, Ian Ranson): --thorough-admission-search, --value-aware-demote, --concurrent-prefill, --recover-invariant-failures.
• Drafting and sampling (Gideon Zenz). --mtp-attention-window N bounds the MTP head's attention to its first 64 keys and the newest N. Post-thinking sampling switches to its own preset once reasoning closes (--post-thinking*, or a post_thinking object per request).
• Serving (Gideon Zenz, Ian Ranson). GET /stats (--stats-port), a dashboard and wedge watchdog in tools/monitor, --request-log-max-mib, --assistant-prefill, --unconstrained-response-format, --lenient-assistant-history, --derive-session-keys, Anthropic ping events, grouped --help, --log-colours, --log-stats-panel, the build id in every binary.
• Vision on CPU and position interpolation (David Oelfke): --vision-residency cpu; --rope-scaling-factor with --rope-scaling-original-context.
• Kernels and conversion (Ian Ranson, Duncan Betts). Programmatic dependent launches in decode graphs (-DNINFER_PDL=ON on compatibility builds), split-KV attention for short prefill steps, a BF16 GEMM fallback, mixed-format MTP banks, the fused RMSNorm with NVFP4 attention input, converters for ModelOpt NVFP4/FP8 and Quasar NVFP4 checkpoints and a least-squares scale search (in grouped_search); a native Windows build against a prebuilt vcpkg tree.
• Unified Linear templates (Neroued). Upstream's Q4/Q5/Q6/Q8, FP8, NVFP4 and BF16 templates and fused projections sit beside this line's routes, and each card takes them only at the widths where they measured faster. Q5 runs 1.5 to 1.7 times as fast from about 8 columns; FP8 and NVFP4 A16 run 1.7 to 7 and 2.5 to 44 times as fast at verify and prefill widths on an RTX 3090. NINFER_LINEAR_ROUTES=legacy|unified forces one table.
• Two-stage GDN prefill (Neroued). Chunks of 16 tokens or more run one preparation pass and one FP32-state recurrence. The op is 1.4 to 4.3 times as fast; Engine prefill of Qwen3.8-27B on an RTX 3090 moved by about 1%. NINFER_GDN_TWO_STAGE=0|1 forces either.
• PackGQA (Gideon Zenz). NINFER_PROMPT_PACK_GQA=1 packs each KV head's query heads into the prompt kernel's tiles: 2.7% faster on an RTX 3090, 0.5% and 3.9% slower on a 4090 and a 5090, so no built-in profile turns it on.
• Engine and serving fixes. Worker out-of-memory recovery (David Oelfke, ported by Ian Ranson); --kv-headroom-mib, --cuda-graph-allowance-mib, --thinking-budget-message (Ian Ranson); --webui-mcp-proxy, table-decoded E8 roots and an SM-count RMSNorm cutoff (tmark00); MTP graph profiles with topology classes (Mykhailo Dementii, upstream PR #221); openable server URLs and CORS preflight echoes (pelebel, natpate). Upstream pull requests: GGUF as a conversion source (giveen), a Q6 recipe (bingchengcc), sparse-MoE, NVFP4 and attention-epilogue tuning (Mykhailo Dementii, Duncan Betts, MOVIBALE), tool-call fixes (Fedor Suchkov, adubkov), Copilot tool shapes (Damian Sromek).

The maintainer map lists each change with the files it touches and the tests that cover it.

Running

Download an artifact from the table below and serve it from the Docker image or from a build. The server speaks the OpenAI and Anthropic APIs, and each card picks up its device profile on its own.

hf download WaveCut/Ternary-Bonsai-2-27B-NInfer-v3 Ternary-Bonsai-2-27B-ninfer-v3.ninfer --local-dir models

# The image: `serve`, the artifact under /models, the flags. Listens on http://localhost:8080/v1.
docker run --rm --gpus all -p 8080:8080 --ulimit memlock=-1 \
  -v "$PWD/models:/models" -v ninfer-cache:/cache \
  ghcr.io/iamwavecut/ninfer-all serve /models/Ternary-Bonsai-2-27B-ninfer-v3.ninfer \
  --model-id bonsai2-27b --max-context 262144 --kv-capacity 262144 --kv-dtype rk8v4 \
  --gdn-state-fp16 --spec dflash2 --draft-tokens 5

# A build: the same flags. Listens on 127.0.0.1:8080 unless --host and --port say otherwise.
ninfer-serve models/Ternary-Bonsai-2-27B-ninfer-v3.ninfer --model-id bonsai2-27b \
  --max-context 262144 --kv-capacity 262144 --kv-dtype rk8v4 --gdn-state-fp16 \
  --spec dflash2 --draft-tokens 5

The recipes below are the configurations the reference tables and the model cards use. They are written for a build; in the container, replace ninfer-serve models/ with the docker run ... serve /models/ line above.

Ternary Bonsai 2 27B: DFlash2, MTP, the largest context, a disk tier

# Fastest single stream: DFlash2 with five drafts over the full 262,144-token window.
ninfer-serve models/Ternary-Bonsai-2-27B-ninfer-v3.ninfer --model-id bonsai2-27b \
  --max-context 262144 --kv-capacity 262144 --kv-dtype rk8v4 --gdn-state-fp16 \
  --spec dflash2 --draft-tokens 5

# MTP drafting through the proposal head.
ninfer-serve models/Ternary-Bonsai-2-27B-ninfer-v3.ninfer --model-id bonsai2-27b \
  --max-context 262144 --kv-capacity 262144 --kv-dtype rk8v4 --gdn-state-fp16 \
  --spec mtp --draft-tokens 3 --lm-head-draft

# The largest context a 24 GB card holds: 958,464 tokens of rk4v4.
ninfer-serve models/Ternary-Bonsai-2-27B-ninfer-v3.ninfer --model-id bonsai2-27b \
  --max-context 958464 --kv-capacity 958464 --kv-dtype rk4v4 --gdn-state-fp16 --rope-yarn

# Adaptive MTP and a disk tier that keeps evicted conversations.
ninfer-serve models/Ternary-Bonsai-2-27B-ninfer-v3.ninfer --model-id bonsai2-27b \
  --max-context 198400 --kv-capacity 198400 --kv-dtype rk8v4 --gdn-state-fp16 \
  --spec mtp --draft-tokens 5 --lm-head-draft --adaptive-mtp \
  --disk-kv-path /var/cache/ninfer --disk-kv-gib 64 --disk-kv-restore

• Draft count. Five drafts are the all-round choice. Seven are faster on short answers; after long documents three to seven win (draft length).
• Images. --vision --vision-residency overlay --vision-max-merged 12288 adds them. The encode borrows the drafter's memory, so the whole window still fits a 24 GB card.
• Past 958,464 tokens. The RTX 5090 and the PRO 6000 hold the 1,048,576-token maximum with rk4v4, DFlash2 or MTP included; the RTX 5090 holds 978,944 with rk8v4. Filled to 1,048,576 tokens, the model misses the code at 90% (about 943K); up to about 880K it found every code on every card.

Qwen3.8-27B: upstream's artifact and the GSQ-RCO IQ3_S release

# Upstream's artifact on a 24 GB card: DFlash2 with five drafts over 245,760 tokens of rk4v4.
ninfer-serve models/qwen3_8_27b.ninfer --model-id qwen3.8-27b \
  --max-context 245760 --kv-capacity 245760 --kv-dtype rk4v4 --gdn-state-fp16 \
  --spec dflash2 --draft-tokens 5

# GSQ-RCO IQ3_S: MTP over 176,128 tokens of rk8v4.
ninfer-serve models/Qwen3.8-27B-GSQ-RCO-IQ3_S-ninfer-v3.ninfer --model-id qwen3.8-27b \
  --max-context 176128 --kv-capacity 176128 --kv-dtype rk8v4 --gdn-state-fp16 \
  --spec mtp --draft-tokens 3

• Upstream's artifact with rk8v4. The same speculation fits 167,936 tokens on an RTX 4090 and 176,128 on an RTX 3090. An RTX 5090 takes the full 262,144 with either KV format.
• IQ3_S. It drafts with --spec dflash2 --draft-tokens 5 as well, and --vision adds images.

Qwen3.8-Flash-Next: experts on the GPUs, in host memory or on disk

hf download WaveCut/Qwen3.8-Flash-Next-GSQ-RCO-Q2_0-NInfer-v3 \
  Qwen3.8-Flash-Next-GSQ-RCO-Q2_0-ninfer-v3.ninfer --local-dir models
hf download WaveCut/Qwen3.8-Flash-Next-ngram-table-NInfer-v3 \
  Qwen3.8-Flash-Next-ngram-table-IQ4_NL-ninfer-v3.ninfer --local-dir models
M=models/Qwen3.8-Flash-Next-GSQ-RCO-Q2_0-ninfer-v3.ninfer
T=models/Qwen3.8-Flash-Next-ngram-table-IQ4_NL-ninfer-v3.ninfer

# Every expert on one RTX PRO 6000 (96 GB).
ninfer-serve $M --ngram-table $T --model-id qwen3.8-flash-next --max-context 32768

# Every expert on two GPUs, one pipeline stage each (Linux): 24 GB cards for Q2_0, 32 GB for IQ3_S.
ninfer-serve $M --ngram-table $T --model-id qwen3.8-flash-next --max-context 32768 --devices 0,1

# One 24 GB GPU: the experts in pinned host memory, the most used of them cached on the GPU.
ninfer-serve $M --ngram-table $T --model-id qwen3.8-flash-next --max-context 32768 --expert-residency host

# One 24 GB GPU and little RAM: the experts stay in the file and stream into a GPU cache.
ninfer-serve $M --ngram-table $T --model-id qwen3.8-flash-next --max-context 32768 --expert-residency disk

# The container, experts in host memory.
docker run --rm --gpus all -p 8080:8080 --ulimit memlock=-1 \
  -v "$PWD/models:/models" -v ninfer-cache:/cache \
  ghcr.io/iamwavecut/ninfer-all serve /models/Qwen3.8-Flash-Next-GSQ-RCO-Q2_0-ninfer-v3.ninfer \
  --ngram-table /models/Qwen3.8-Flash-Next-ngram-table-IQ4_NL-ninfer-v3.ninfer \
  --model-id qwen3.8-flash-next --max-context 32768 --expert-residency host

• Other releases. IQ3_S and the Coder IQ1_M build take the same flags with their own file and the same table. The tables above show the placements measured for each.
• Memory. Host experts pin 34 GB of RAM for Q2_0, 50 GB for IQ3_S and 25 GB for the Coder build. Disk experts need under 1 GB of RAM; the page cache does the rest. The GPU expert cache takes what is free after startup, or --expert-cache-mib.
• The n-gram table. Its rows are read from the file, 16 per token, unless --ngram-ram loads all 28.8 GB. A model started without its table is refused; --no-ngram-table overrides that, an experimental mode with no practical use.
• Images and video. --vision adds the Vision tower (0.9 GB on the GPU).

Qwen3.6-35B-A3B NVFP4: RTX 50 series and RTX PRO 6000 only

ninfer-serve models/Qwen3.6-35B-A3B-NVFP4-ninfer-v3.ninfer --model-id qwen3.6-35b-a3b \
  --max-context 262144 --kv-capacity 262144 --kv-dtype rk8v4 --gdn-state-fp16 \
  --spec mtp --draft-tokens 3 --lm-head-draft --vision

The NVFP4 experts prefill through W4A4 on Blackwell's FP4 tensor cores, so this needs a 120a build (the image has one); sm_86 and sm_89 builds refuse the file. On an RTX 5090 it starts in 17 s and leaves 7.4 GiB of the card free.

Qwen3.8-27B fine-tunes: Huihui abliterated, HauhauCS Aggressive, MXFP8-CRACK

All three have the official Qwen3.8-27B identity and size, so the Qwen3.8-27B recipes above serve them too; CRACK has no DFlash2 adapter. Their cards' configurations:

# Huihui abliterated on one RTX 3090: 198,400 tokens with MTP and Vision.
ninfer-serve models/Huihui-Qwen3.8-27B-abliterated-ninfer-v3.ninfer --model-id qwen3.8-27b \
  --max-context 198400 --kv-capacity 198400 --kv-dtype rk8v4 --gdn-state-fp16 \
  --spec mtp --draft-tokens 3 --lm-head-draft \
  --vision --vision-residency overlay --vision-max-merged 12288

# HauhauCS Aggressive: DFlash2 with seven drafts.
ninfer-serve models/Qwen3.8-27B-Uncensored-HauhauCS-Aggressive-DFlash2-ninfer-v3.ninfer \
  --model-id qwen3.8-27b --max-context 32768 --kv-capacity auto \
  --spec dflash2 --draft-tokens 7 --lm-head-draft

# MXFP8-CRACK: MTP.
ninfer-serve models/Qwen3.8-27B-MXFP8-CRACK-ninfer-v3.ninfer --model-id qwen3.8-27b \
  --max-context 32768 --kv-capacity auto --spec mtp --draft-tokens 3 --lm-head-draft

A card without a built-in profile

ninfer-calibrate --print > my-gpu.json

The engine measures it by itself at first start. Running it by hand refreshes the profile after a driver or clock change. See device profiles.

Artifacts

The Bonsai, GSQ-RCO and Flash-Next artifacts use formats that only this line reads.

model	artifact	size	contents
Ternary Bonsai 2 27B	WaveCut/Ternary-Bonsai-2-27B-NInfer-v3	8.87 GiB	ternary weights, Vision, Bonsai-trained MTP head and DFlash2 adapter, exact proposal head
Qwen3.8-27B	neroued/Qwen3.8-27B-NInfer	19.03 GiB	upstream's groupwise-int (Q4/Q5) with MTP and DFlash2; the reference tables' artifact
Qwen3.8-27B GSQ-RCO IQ3_S	WaveCut/Qwen3.8-27B-GSQ-RCO-IQ3_S-NInfer-v3	13.99 GiB	ISTA-DASLab's 3.5-bit GGUF blocks byte for byte, Q6_K MTP head, Vision, DFlash2, proposal head
Qwen3.8-27B, abliterated	WaveCut/Huihui-Qwen3.8-27B-abliterated-NInfer-v3	19.03 GiB	upstream's qwen3_8_27b recipe with MTP, DFlash2 and a proposal head
Qwen3.8-27B Uncensored, HauhauCS Aggressive	WaveCut/Qwen3.8-27B-Uncensored-HauhauCS-Aggressive-DFlash2-NInfer-v3	19.03 GiB	groupwise-int with the tune's MTP head, DFlash2 and a proposal head
Qwen3.8-27B MXFP8-CRACK	WaveCut/Qwen3.8-27B-MXFP8-CRACK-NInfer-v3	16.96 GiB	groupwise-int with MTP and a proposal head; no DFlash2
Qwen3.8-Flash-Next GSQ-RCO Q2_0	WaveCut/Qwen3.8-Flash-Next-GSQ-RCO-Q2_0-NInfer-v3	35.89 GiB	ISTA-DASLab's 2.4-bit GGUF blocks byte for byte, Vision; reads the n-gram table
Qwen3.8-Flash-Next GSQ-RCO IQ3_S	WaveCut/Qwen3.8-Flash-Next-GSQ-RCO-IQ3_S-NInfer-v3	51.90 GiB	the 3.5-bit release, the same way
Qwen3.8-Flash-Next Coder GSQ-RCO IQ1_M	WaveCut/Qwen3.8-Flash-Next-GSQ-RCO-Coder-IQ1_M-NInfer-v3	28.42 GiB	the expert-pruned coding build (256 experts per layer), the same way
Qwen3.8-Flash-Next n-gram table	WaveCut/Qwen3.8-Flash-Next-ngram-table-NInfer-v3	26.82 GiB	the IQ4_NL n-gram table every Flash-Next artifact reads (--ngram-table)
Qwen3.6-35B-A3B NVFP4	WaveCut/Qwen3.6-35B-A3B-NVFP4-NInfer-v3	20.39 GiB	RedHatAI's NVFP4 experts code for code, Q8 projections, Vision, MTP, proposal head; sm_120a GPUs only

The official NInfer artifacts listed in the original READMEs load here too. Weight conversion shows how the Bonsai and GSQ-RCO artifacts are built.

Docker

ghcr.io/iamwavecut/ninfer-all:latest is built from every master commit that passes CI, on CUDA 13.4 and Ubuntu 26.04. It carries two builds and starts the one that matches the GPU: sm_86 for the RTX 30 and RTX 40 series, sm_120a for the RTX 50 series and the RTX PRO 6000 Blackwell.

• Host. An NVIDIA driver of the CUDA 13 branch (580 or newer) and the NVIDIA Container Toolkit.
• Tags. latest, sha-<commit> and the VERSION; the image is 1.6 GB compressed.
• Volumes. /models holds artifacts, /cache the device profile measured on first start, and /grafts/<model>/ optional prompt grafts.
• Checked on an RTX 3090 and an RTX 5090 with driver 580.159.03.

The container's command chooses what runs:

command	runs
serve, ninfer, perplexity, calibrate [args]	that binary with any artifact and flags; serve listens on 0.0.0.0:8080 unless given --host or --port
run <model> [profile] (the default: run qwen38-27b)	the launcher profiles of scripts/run.sh for upstream's qwen38-27b and qwen36-35b-a3b, with its NINFER_* overrides (-e NINFER_SPEC=mtp, -e NINFER_CONTEXT=131072, ...)
download <model>	scripts/download-model.sh into /models: qwen38-27b, qwen36-27b or qwen36-35b-a3b

Upstream's Qwen3.8-27B with the measured tuned profile, on http://localhost:8080/v1:

docker run --rm -v "$PWD/models:/models" ghcr.io/iamwavecut/ninfer-all download qwen38-27b
docker run --rm --gpus all -p 8080:8080 --ulimit memlock=-1 \
  -v "$PWD/models:/models" -v ninfer-cache:/cache \
  ghcr.io/iamwavecut/ninfer-all run qwen38-27b

compose.yaml wires the GPU, the port and the volumes for the same: docker compose run --rm ninfer download qwen38-27b, then docker compose up -d. NINFER_IMAGE_ARCH=sm86|sm120a overrides the GPU detection, and docker build -t ninfer . builds the image from source (--build-arg ARCHS=86 for one architecture).

Building

Linux with CUDA 13.1

cmake -S . -B build -G Ninja -DCMAKE_BUILD_TYPE=Release -DCMAKE_CUDA_ARCHITECTURES=86
cmake --build build --target ninfer-serve ninfer-calibrate

CMAKE_CUDA_ARCHITECTURES is 86 for the RTX 30 series, 89 for the RTX 40 series and 120a for the RTX 50 series and the RTX PRO 6000 Blackwell (on the mma.sync compatibility path, which the ternary route needs). A 120a build needs CUDA 13.1 or newer: CUDA 12.8 and 12.9 miscompile sm_120a kernels, and configure refuses them. The opt-in build options are listed in the Linux build guide. Windows builds, release packages, tests and benchmarks work as in the NInfer-3090 README.

License

Apache-2.0, as upstream. The Bonsai artifact's weights come from PrismML, ProCreations and Qwen, all Apache-2.0; its card lists the notices. The Qwen3.8-Flash-Next artifacts carry the Qwen Community License 1.0 of their model.