[IMPORTANT] [Bionic/MLX] Context auto-fits to 4,864 but Bionic still rejects requests as too large

[imCinq]

Which version of LM Studio?
LM Studio: 0.4.20+1
Bionic: 1.0.3+3
CLI commit: 71bd99c

Which operating system?
macOS 26.5.2 — build 25F84
MacBook Air, Apple M2
Unified memory: 16 GB

What is the bug?
When loading prism-ml/Ternary-Bonsai-27B-mlx-2bit through Bionic/LM Studio, the requested context length is automatically reduced by LM Studio from 6,144 to 4,864 tokens because of the memory guardrail.

The loaded instance correctly shows:

Ternary Bonsai 27B MLX 2bit 4864 ctx Loaded

However, Bionic still reports:

The selected model was loaded with a context length too small for this request.

Changing the requested context length to 5,000, 6,144, and later 4,096 did not resolve the error. The model remains loaded and idle, but Bionic cannot issue a request.

This appears to be a mismatch between the requested/configured context length and the actual auto-fitted context length reported by the loaded model instance.

Screenshots

Logs
At a requested context of 6,144:

text [2026-07-27 17:14:31][DEBUG][context_fit][INFO]: Model context auto-fit: family=qwen3_5 max=262,144 fitted=4,864 safe_ceiling=8.84GiB estimated_peak=8.84GiB

[2026-07-27 17:14:31][DEBUG][cache_store][INFO]: VLM prompt cache context target: configured=6,144 fitted=4,864 effective=6,144

The loaded-instance UI and lms ps both report 4,864, despite the cache-store entry retaining effective=6,144.

LM Studio’s memory estimate also reports that the 6,144-token configuration cannot be loaded under its current guardrails:

text canLoad=false estimatedUsageBytes=11929518865.599998 estimatedModelUsageBytes=8521084904 estimatedContextUsageBytes=3408433961.5999994

After changing the requested context to 4,096:

text [2026-07-27 17:29:14][DEBUG][context_fit][INFO]: Model context auto-fit: fitted=4,864 safe_ceiling=8.84GiB estimated_peak=8.84GiB

[cache_store][INFO]: VLM prompt cache context target: configured=4,096 fitted=4,864 effective=4,864

To Reproduce
Steps to reproduce the behavior:
Use a Mac with Apple Silicon and 16 GB unified memory.

Install LM Studio 0.4.20+1 and Bionic 1.0.3+3.

Download and select:
prism-ml/Ternary-Bonsai-27B-mlx-2bit
Load the model through Bionic with a requested context of 6,144.

Observe that the loaded instance reports 4,864 ctx.

Send a Bionic request.

Observe the context-too-small error.

Eject and reload with a requested context of 4,096.

Confirm that LM Studio still reports an actual context of 4,864.

Retry the request; the same error persists.

[imCinq]

WHAT HAS BEEN TRIED:
INCREASING/DECREASING CONTEXT
LIMITING RESPONSE CONTEXT (512/1000)
LETTING A MODEL TRY TO FORCE BIONIC TO CHANGE CONTEXT

[M3tu20222]

same problem on me
with MLX models lmstudio fails to accept the context number.
one month ago alla mlx models were accepting 40K context, but now lmstudio doesnt accept context size

---

cold run log:

[context_fit][INFO]: Model context auto-fit: family=gemma4 max=262,144 fitted=4,352 working_set=17.76GiB reserve=3.00GiB safe_ceiling=14.76GiB baseline=13.82GiB full_kv=20480B/token prompt_inputs=5632B/token attention=65536B/token rotating_peak=0.59GiB fixed_ssm=0.00GiB estimated_peak=14.78GiB
[cache_store][INFO]: VLM prompt cache context target: configured=8,192 fitted=4,352 effective=8,192

[gmancode]

Bionic

model
llm qwen/qwen3.5-35b-a3b

Error
The selected model was loaded with a context length that is too small for this request.

Try one of the following:
• Reduce load on your machine (close apps you're not using) and re-load the model.
• Start with a shorter request or reduce attachment size.

[p3ngwin]

same problem here.

i was running Gemma 4 12B QAT, and Gemma 4 31B QAT, fine for most of the night, then got this strange error.

exiting LM Studio Bionic, completely shutting it down, and it components (llama-server.exe) with Win-task manager, then reloading, results in same problem.

App version
1.0.4 (build 9)
Channel
Stable

[Alexander3847575]

It refuses to generate only when I have overflow context set to truncate in middle. Works fine with rolling context window. However, it will still use the auto fit size.

[JetWest]

Adding another Windows + NVIDIA data point:

Environment

Windows 11 Pro, HP OMEN MAX 45L
RTX 5090, 32GB VRAM; 32GB DDR5
LM Studio + Bionic (1.0.x)
Model: MLX q4 17B build (qwen/qwen3.8-17b-17b-a3b-mlx-q4, file ends .mlx)
What happened

The Context Length settings box showed my requested 65,536, but the loaded-model badge read "4k" (and later "16k" after a reload) — the auto-fit silently shrank the context far below the requested value, with no warning.
The "reload the page" prompt at the top of the page did nothing.
Every Bionic chat then froze right after a tool call, with the exact error from the OP: "The selected model was loaded with a context length that is too small for this request."
lms ps showed the true loaded context — it is the source of truth; the settings box was showing the requested value, not the actual one.
Compounding factor: opening an old Bionic chat auto-reloads that chat's pinned model, so the broken model reappeared even after a working model was loaded.
Workaround that worked for me

lms unload
lms load qwen/qwen3.5-9b-gguf --context-length 65536
lms ps -> "9.7GiB, 64k ctx" OK
The GGUF / llama.cpp path honored the configured context; the MLX path did not.

Happy to provide logs or more detail.